NBER WORKING PAPER SERIES
TECHNOLOGY ENTRY IN THE PRESENCE OF PATENT THICKETS
Bronwyn H. HallChristian Helmers
Georg von Graevenitz
Working Paper 21455http://www.nber.org/papers/w21455
NATIONAL BUREAU OF ECONOMIC RESEARCH1050 Massachusetts Avenue
Cambridge, MA 02138August 2015
We are grateful to the Intellectual Property Office of the United Kingdom for research support and to the National Institute of Economic and Social Research for hospitality while this paper was being written. The views expressed here are those of the authors. They are not necessarily those of the UK Intellectual Property Office or NIESR. The revision of this paper has benefitted from comments by participants in seminars at the UKIPO, the USPTO, the NBER Summer Institute, the University of California at Berkeley, TILEC, Tilburg University, the University of Wuppertal and the TIGER Conference in Toulouse. We would like to thank Jonathan Haskel, Scott Stern, Dietmar Harhoff, Stefan Wagner and Megan MacGarvie for comments, and Iain Cockburn, Dietmar Harhoff, and Stefan Wagner for very generous support with additional data. The views expressed herein are those of the authors and do not necessarily reflect the views of the National Bureau of Economic Research.
At least one co-author has disclosed a financial relationship of potential relevance for this research. Further information is available online at http://www.nber.org/papers/w21455.ack
NBER working papers are circulated for discussion and comment purposes. They have not been peer-reviewed or been subject to the review by the NBER Board of Directors that accompanies official NBER publications.
© 2015 by Bronwyn H. Hall, Christian Helmers, and Georg von Graevenitz. All rights reserved. Short sections of text, not to exceed two paragraphs, may be quoted without explicit permission provided that full credit, including © notice, is given to the source.
Technology Entry in the Presence of Patent ThicketsBronwyn H. Hall, Christian Helmers, and Georg von Graevenitz NBER Working Paper No. 21455August 2015, Revised January 2017JEL No. K11,L20,O31,O34
ABSTRACT
We analyze the effect of patent thickets on entry into technology areas by firms in the UK. We present a model that describes incentives to enter technology areas characterized by varying technological opportunity, complexity, and the potential for hold-up due to the presence of patent thickets. We show empirically that our measure of patent thickets is associated with a reduction of first time patenting in a given technology area controlling for the level of technological complexity and opportunity. Technological areas characterized by more technological complexity and opportunity, in contrast, see more entry. Our evidence indicates that patent thickets raise entry costs, which leads to less entry into technologies regardless of a firm’s size.
Bronwyn H. Hall123 Tamalpais RoadBerkeley, CA 94708and [email protected]
Christian HelmersDepartment of EconomicsSanta Clara UniversityLucas HallSanta Clara, CA [email protected]
Georg von GraevenitzQueen Mary University of [email protected]
1
1 Introduction
The past two decades have seen an enormous increase in patent filings worldwide (Fink
et al., 2013). There are signs that the high level of patenting may be reducing innovation
in certain technologies (Federal Trade Commission, 2011; Bessen and Meurer, 2008;
Jaffe and Lerner, 2004; Federal Trade Commission, 2003). Companies drawing on these
technologies face elevated legal costs of commercializing innovative products when
patents that contain overlapping claims form so-called “patent thickets” (Shapiro,
2001). Patent thickets arise where individual products draw on innovations protected
by hundreds or even thousands of patents, often with fuzzy boundaries. These patents
belong to many independent and frequently competing firms. Patent thickets can lead to
hold-up of innovations, increases in the complexity of negotiations over licenses,
increases in litigation, and they create incentives to add more, often weaker patents to
the patent system (Allison et al., 2015). The increased transaction costs associated with
patent thickets reduce profits that derive from the commercialization of innovation, and
ultimately may reduce incentives to innovate.
There is a growing theoretical (Bessen and Maskin, 2009; Clark and Konrad, 2008;
Farrell and Shapiro, 2008; Fershtman and Kamien, 1992) and legal literature (Chien and
Lemley, 2012; Bessen et al., 2011) on patent thickets. Related work analyzes firms’
attempts to form patent pools to reduce hold-up (Joshi and Nerkar, 2011; Lerner et al.,
2007; Lerner and Tirole, 2004) and the particular challenges posed in this context by
standard essential patents (Lerner and Tirole, 2013).
The existing empirical evidence on patent thickets is largely concerned with showing
that they exist and measuring their density (von Graevenitz et al., 2011; Ziedonis, 2004).
There is less evidence on the effects patent thickets have for firms. Cockburn and
MacGarvie (2011) demonstrate that patenting levels affect product market entry in the
software industry. They show that a 1 per cent increase in the number of existing
patents is associated with a 0.8 per cent drop in the number of product market entrants.
This result echoes earlier findings by Lerner (1995) who showed for a small sample of
U.S. biotech companies that first-time patenting in a given technology is affected by the
presence of other companies’ patents. These two papers use patent counts in narrow
technological fields as a proxy for thickets; in this paper we move beyond this measure
to use an indicator of the extent to which the patents actually might overlap.
Bessen and Meurer (2013) suggest that patent thickets also lead to increased litigation
related to hold-up. They (and we) use the term to describe a situation where an alleged
infringer faces the threat of an injunction or high licensing costs after she has sunk
investment.1 Patent thickets have remained a concern of antitrust agencies and
1 “High licensing costs” refers to costs that are higher than those that would have been negotiated ex ante
in the present of possible “invent around” before the alleged infringer sank her investment. This
2
regulators in the U.S. for over a decade (Federal Trade Commission, 2011, 2003; USDoJ
and FTC, 2007). Reforms that address some of the factors contributing to the growth of
patent thickets have recently been introduced in the U.S. (America Invents Act (AIA) of
2011) 2 and by the European Patent Office (EPO).
However, it is sometimes argued that patent thickets are a feature of rapidly developing
technologies in which technological opportunities abound (Teece et al., 2014). Thickets
are seen as a reflection of fast technological progress that is paired with increased
technological complexity (Lewis and Mott, 2013). This suggests that the increased
transactions costs associated with patent thickets and the benefits of technological
complexity and opportunity often coincide. There may be a trade-off between
technological opportunity and growth on the one hand and increased transaction costs
due to the emergence of patent thickets on the other - if the transaction costs of
patenting in complex technologies are not avoidable. The challenge in assessing
technologies with high levels of patenting is to develop a framework that captures the
main factors that incentivize patenting and the costs and benefits thereof.
This paper contains such a framework in the form of a model that incorporates
technological patenting incentives and the effects of patenting by rivals on each firm’s
incentives to enter and to patent. The aim of the paper is to empirically disentangle the
effects of thickets and technological opportunity and complexity on entry into
technology areas. Modeling patenting and entry we show that greater technological
opportunity and complexity both encourage entry, while higher transaction costs
associated with preexisting patent thickets reduce entry. These effects are quantified
empirically using firm-level data on patenting by firms in the United Kingdom.
In a first contribution, we develop a theoretical model of entry into patenting in discrete
and complex technologies. Building on previous work on patenting incentives (von
Graevenitz et al., 2013), we model how entry decisions are affected by the potential for
hold-up due to patent thickets. The model focuses on the interaction between
incumbents and new entrants through two channels: i) via legal costs associated with
patent enforcement and ii) via incumbency advantages in R&D fixed costs. In contrast to
previous work, we separate technological complexity and thicket density. Where
previously granted patents raise thicket density, we assume that the expected legal
costs of defending new patents are higher and the benefits of amassing portfolios of
new patents are lower. This reflects higher potential for hold-up and leads firms to
patent less and reduce entry, ceteris paribus. As in von Graevenitz et al. (2013), the
model predicts that greater complexity of a technology leads to an increase of a firm’s
possibility can arise because of either prohibitive search costs (Mulligan and Lee, 2012) or fuzzy patent
boundaries or both.
2 For further information see http://www.gpo.gov/fdsys/pkg/BILLS-112hr1249enr/pdf/BILLS-112hr1249enr.pdf
3
patent applications, while greater technological opportunity reduces applications.3 The
model also predicts that higher complexity and opportunity will increase entry, because
competition for individual patents is reduced and the probability that entrants can
establish themselves in a technology is increased. In addition, where incumbency
implies that firms enjoy lower costs of R&D, incumbents patent more than entrants,
which reinforces the growth of thickets.
Our second contribution consists of an empirical test of these predictions and a
quantification of the importance of the conflicting effects of opportunity, complexity,
and patent thickets on entry. We use patent data from both the EPO and the USPTO to
construct measures of technological opportunity, technological complexity and hold-up
potential and relate these to entry into new technology areas by UK firms.
To test the predictions of the model, we measure technological complexity and hold-up
potential in patent thickets separately. This is an important improvement over the
analysis in von Graevenitz et al. (2013), who conflated complexity and hold-up potential
arising from existing patent portfolios. To distinguish complexity and hold-up potential,
we introduce a new empirical measure of technological complexity and sharpen the
definition of an existing measure of hold-up potential.4
Our empirical analysis of entry confirms that greater technological opportunity and
complexity increase entry and that hold-up potential reduces entry substantially. While
we cannot quantify the overall net welfare effect, our results indicate that patent
thickets raise entry costs for large and small firms alike. To the extent that more original
and radical, rather than incremental ideas come from new entrants rather than
incumbents (Tushman and Anderson, 1986; Henderson, 1993), this is likely to have
negative long-run consequences on innovation and product market competition. In
combination with earlier results by von Graevenitz et al. (2013), who point to a positive
correlation between patenting levels and the presence of thickets, our results suggest
that any increases in transaction costs due to thickets can potentially have important
dynamic effects on innovation by reducing entry.
The remainder of this paper is organized as follows. Section 2 presents a model of entry
into patenting in a technology area and derives several testable predictions. Section 3
describes the data, and the empirical measurement of the key concepts in the model.
Section 4 discusses our results and Section 5 provides concluding remarks.
3 Opposite signs of complexity and opportunity are due to firms’ efforts to maintain a constant share of
patents on each opportunity, to protect their profits from patenting (von Graevenitz et al., 2013).
4 We also provide separate empirical tests of the measures we propose on the EPO patent data used by
von Graevenitz et al. (2013) in Appendix D.
4
2 Theoretical Model
This section summarizes the main results of a two-stage model of entry into patenting
and subsequent patenting decisions. The model shows how complexity of a technology,
technological opportunity and the potential of hold-up due to patent thickets affect
firms’ decisions to enter into patenting in that technology.
We assume that each technology consists of a set of opportunities, each of which
consists of a number of patentable facets. The more facets opportunities in a technology
have, the more complex is that technology.
Greater technological opportunity attracts entrants to a technology as more avenues
arise to earn a profit through application of the technology. Greater complexity of a
technology also attracts entrants, because it is more likely that each entrant will be able
to gain a share of the profits flowing from any one opportunity. Where multiple firms
hold patents on the same opportunity, licensing negotiations or litigation ensue as firms
divide the profits flowing from the opportunity. We assume that holding a larger share
of patents on an opportunity is beneficial for firms in terms of licensing or litigation, but
less so when thickets arise from poorly delineated patents that provide increased
options to litigate. This captures the costs imposed by thickets on patentees.
The model consists of two stages: entry and patenting. We solve the model by backward
induction. Analysis of the entry decision shows that greater technological opportunity
and complexity increase entry into patenting, while the threat of increased legal costs
arising from denser patent thickets reduces entry into patenting.
Analyzing patenting we show that denser thickets reduce patent applications, while
greater technological opportunity increases applications in discrete technologies, but
reduces them in complex technologies. This prediction is analogous to that in von
Graevenitz et al. (2013). Finally, greater complexity increases patent applications.
Below we set out the notation and the main findings from the model. Most proofs and
derivations are relegated to appendices A and B.
2.1 Notation and Assumptions
The key variables of the model are the complexity of a technology k, measured by Fk
(𝐹𝑘 ∈ ℝ0+), the degree of technological opportunity, measured by Ok (𝑂𝑘 ∈ ℝ0
+), and
hold-up potential hk. The value of all ��𝑘 patents in an opportunity is Vk. In the simplest
discrete setting this is the value of the one patent (facet) that covers each technological
opportunity. In more complex technologies this is the value of controlling all patents
(facets) on a technological opportunity. Firms (indexed by i) choose the number of
opportunities oi to invest in and the number of facets fi per opportunity to patent.
In equilibrium only 0 1ˆ(1 (1 ( / ) ))N
k k kF f F
facets are patented, where kf is the
equilibrium number of facets chosen by applicants and NO is the number of firms that
5
have chosen a specific opportunity.5 Since may be smaller than Fk the total value of
patenting in a technology is . Von Graevenitz et al. (2013) assume that the
value function is convex in covered facets. In Appendix B.1 we discuss how this
assumption is relaxed and generalize the model by introducing a concave function
relating the share of patents a firm holds on an opportunity sik to the proportion of the
value Vk the firm can extract through licensing and its own sales: Δ(sik). This captures
the benefits that a patent portfolio confers in the market for technology.
In sum, the assumptions we make on the value function and portfolio benefits are:
( ): (0) 0, 0k
VVF V
F
(1)
2
2
( ) ( )( ): (0) 0, 0 and 0ik ik
ik ik
d s d sPB
ds ds
(2)
The model contains three types of patenting costs:
• Costs of R&D per opportunity, which depend on overall R&D activity in that
technology area: .
• Costs of maintaining each granted patent in force: Ca.
• Costs of coordinating R&D on different technological opportunities Cc(oi), where
0c
i
C
o
(3)
These assumptions imply that R&D costs are fixed costs.6 We allow for the endogenous
determination of the level of R&D fixed costs, which rise as more opportunities are
researched simultaneously by rival firms. This reflects competition for inputs into R&D,
e.g. scientists and engineers that are in fixed supply in the short run (Goolsbee, 1998).
Where multiple firms own facets on an opportunity, their legal costs L(γik, sik, hk) depend
on the absolute number of patented facets γik, on the share of patents per opportunity
that a firm holds sik, and on the extent to which they face hold-up hk. The first two
channels capture the costs of defending a patent portfolio as the number of patents
increases, while leaving scope for effects on bargaining costs that derive from the share
5 The properties of N0 are summarized in Appendix A.2.
6 It also implies that there is no technological uncertainty. However, introducing technological
uncertainty into the model does not change the main comparative statics results.
kF
( ) ( )k kV F V F
( )k kV F
0ON
jjC o
6
of patents owned: The hold-up parameter captures contexts in which several firms’ core
technologies become extremely closely intertwined. Then each firm has to
simultaneously negotiate with many others to commercialize its products, which
significantly raises transaction costs.
2 2
2 2
2 2
( ): ( , , ), where 0, 0, 0, 0,
0, 0, 0
ik ik k
ik ik ik ik
k ik k ik k
L L L LLC L s h
s s
L L L
h h s h
(4)
All remaining cross partial derivatives of the legal costs function are zero.
In what follows, we use the following definitions:
( ) ( ), , , , and ( ) ( )
i i k k k k i kk k k k k
k k k k k k k i
o f F V F s d s f F
O F V F F s ds F f
.
Here 𝜔𝑘 is the share of opportunities each firm chooses to pursue, 𝜙𝑘 is the share of
facets each firm seeks to patent per opportunity, 𝜇𝑘 is the elasticity of the value function
w.r.t. the level of complexity, 𝜉𝑘 is the elasticity of the benefits function ∆ w.r.t. the share
of patents each firm is granted and 𝜂𝑘 is the elasticity of the number of covered facets
w.r.t. the number of patent applications of each firm.
2.2 Patenting and Entry
Firm i’s profits in technology k, , are a function of the number of
opportunities oi which the firm invests in, the number of facets per opportunity fi the
firm seeks to patent, the total number of patentable facets per opportunity Fk, the
number of technological opportunities a technology offers Ok, the number of firms
entering the technology Nk, and the degree of hold-up in that technology hk.
In this section we analyze the following two-stage game G*:
Stage 1: Firms enter until ( , , , , , ) 0ik i i k k k ko f F O N h ;7
Stage 2: Firms simultaneously choose the number of opportunities, oi, to invest in and
the number of facets per opportunity fi to patent in order to maximize profits πik.
We solve the game by backward induction and derive local comparative statics results
for the symmetric extremal equilibria of the second stage game. For the subsequent
analysis it is important to note that all equilibria of this second stage game are
symmetric. In case that the second stage game has multiple equilibria we focus on the
7 We treat Nk as a continuous variable here, which is an abstraction that simplifies our analysis.
( , , , , , )ik i i k k k ko f F O N h
7
properties of the extremal equilibria when providing comparative statics results
(Milgrom and Roberts, 1994; Amir and Lambson, 2000; Vives, 2005). Equilibrium
values of the firms’ choices are denoted by a hat (^) and we drop the firm specific
subscripts in what follows, e.g., .
At stage two of the game each firm maximizes the following objective function:
0( , ) ( ) ( ) ( , , ) ( ) ( )ON
ik i i i k ik ik ik k j i k a c ijo f o V F s L s h C o f p C C o (5)
This expression shows that per opportunity k, the firm derives profits from its share
of patented facets, while facing legal costs L to appropriate those profits,
as well as costs of R&D C0, costs of maintaining its patent portfolio Ca, and coordination
costs across opportunities Cc.
2.3 Simultaneous Entry with Multiple Facets
2.3.1 Comparative statics of patenting
We show that the second stage of this game is smooth supermodular:
Proposition 1: The second stage patenting game, defined in particular by assumptions
(VF, eq. 1), (PB, eq. 2) and (LC, eq. 4) is smooth supermodular if k ik and if ownership
of the technology is expected to be fragmented.
This is shown in Appendix B.1.
This result generalizes Proposition 1 derived by von Graevenitz et al. (2013).8 Given this
result we can show that:
Proposition 2: The potential for hold-up in complex technologies reduces patenting
incentives.
In Appendix B.2 we show that the expected legal costs of hold-up reduce the number of
opportunities that firms invest in. In addition, firms with larger portfolios are more
exposed to hold-up and benefit less from the share of patents they have patented per
opportunity. Both effects combine to reduce the number of facets each firm applies for.
2.3.2 Comparative statics of entry
In Appendix B.3 we show that there is a free entry equilibrium. In this equilibrium the
following propositions hold:
Proposition 4: Under free entry greater complexity of a technology increases entry.
8 Here it is no longer the case that the value function has to be increasing in the number of patented facets
for supermodularity of the patenting game. We relegate further discussion of this result to Appendix B.1.
k
/ik k i ks p f F
8
In the model, complexity has countervailing effects: first, it increases profits, because it
is less likely that duplicative R&D arises making each opportunity more valuable; this
clearly increases incentives to enter. Next, given the level of patent applications 𝑓𝑘,
complexity reduces the probability that each facet is patented, which reduces profits
and entry incentives. Finally, complexity reduces competition for each facet, which
increases the probability of patenting and increases innovation incentives. Overall we
show that the positive effects outweigh the negative effects and incentives for entry rise
with complexity of a technology.
To derive Proposition 4, consider how equilibrium profits are affected by the
complexity of the technology Fk, the degree of technological opportunity Ok, and the
potential for hold-up hk:
𝜕𝜋(��,��)
𝜕𝐹𝑘= ��
��𝑘
𝐹𝑘((𝜀��𝑘,𝐹𝑘
− 𝜀��𝑘,𝐹𝑘��𝑘) [𝑉 (��𝑘)
△(��𝑘)
��𝑘(��𝑘 − 𝜉𝑘) +
𝜕𝐿
𝜕��𝑘]) > 0 (6)
𝜕𝜋(��,��)
𝜕𝑂𝑘= ��
𝜕��𝑜
𝜕𝑂𝑘
��𝑘
��𝑜((𝜀��𝑘,𝑁𝑜
− 𝜀��𝑘,𝑁𝑜��𝑘) [𝑉 (��𝑘)
△(��𝑘)
��𝑘(��𝑘 − 𝜉𝑘) +
𝜕𝐿
𝜕��𝑘] −
𝜕𝐶𝑜
𝜕��𝑜��
��𝑜��
𝜕��𝑘) > 0 (7)
𝜕𝜋(��,��)
𝜕ℎ𝑘= −��
𝜕𝐿
𝜕ℎ𝑘< 0 (8)
Proposition 4 follows from the Implicit Function theorem once we know the sign of the
derivative of profits w.r.t. F. Under free entry firms’ profits decrease with entry:
𝜕𝑁𝑘
𝜕𝐹𝑘= −
𝜕𝜋
𝜕𝐹𝑘
𝜕𝜋
𝜕𝑁𝑘⁄ (9)
Therefore, the Implicit Function theorem implies that the sign of the effect of
complexity F on entry depends on the sign of the effect of complexity on profits.
Equation (6) shows that the effect of complexity on profits depends on the difference
between the elasticities 𝜀��𝑘,𝐹𝑘 and 𝜀��𝑘,𝐹𝑘
��𝑘, which are derived in Appendix A.3. The
elasticity 𝜀��𝑘,𝐹𝑘 is (see Appendix A.1):
𝜀��𝑘,𝐹𝑘= ��𝑜
2��𝑘−
1
2(1+
1
��𝑜)
1−��𝑘 (10)
This elasticity is negative for ��𝑘 <1
2, which is also a precondition for supermodularity of
game G*. Both terms in brackets in equation (6) are positive, when game G* is
9
supermodular. This implies that greater complexity raises profits and this induces
entry.9
Proposition 5: Under free entry greater technological opportunity increases entry.
For any given number of entrants an increase in technological opportunity reduces
competition between firms for patents. This increases firms’ expected profits and
increases entry.
Continuing from the proof of Proposition 4 above, by the Implicit Function theorem the
sign of the derivative of profits w.r.t. technological opportunity determines the effect of
technological opportunity on entry:
𝜕𝑁𝑘
𝜕𝑂𝑘= −
𝜕𝜋
𝜕𝑂𝑘
𝜕𝜋
𝜕𝑁𝑘⁄ (11)
An increase in technological opportunity increases profits and entry. In Appendix B.3
we show that the term in brackets in Equation (7) is negative under free entry. Profits
increase as technological opportunity increases, because fewer firms enter per
opportunity.
Proposition 6: Under free entry the potential for hold-up reduces entry.
An increase in the potential for hold-up raises firms’ expected legal costs. This reduces
expected profits and lowers potential for entry.
To derive this prediction, note that by the Implicit Function theorem the sign of the
derivative of profits w.r.t. the level of hold-up in a technology area determines the effect
of hold-up on entry:
𝜕𝑁𝑘
𝜕ℎ𝑘= −
𝜕𝜋
𝜕ℎ𝑘
𝜕𝜋
𝜕𝑁𝑘⁄ (12)
Hence, equation (8) shows that the effect of hold-up on entry derives from the increased
legal costs that the possibility of hold-up imposes on affected firms.
2.4 Entry and Incumbency
The previous section sets out a model in which all firms entered and then invested in
patents. At both stages firms’ decisions were simultaneous. In Appendix B.5 we extend
the model to a setting in which some firms, the incumbents, face lower costs ( ,
9 When ��𝑘 ≥1
2 we no longer have the assumptions necessary to show supermodularity. This situation
corresponds to the case where one firm has more than half the patents in a particular technology
opportunity within a technology area. Thus our results may not hold when a specific opportunity is highly
concentrated. In general this will not be the case, especially at our level of empirical analysis, but it would
be interesting to explore this possibility in future work.
OC
10
where ) of entering opportunities. This captures the fact that incumbents have
previous experience of doing R&D in a technology area. We demonstrate that our main
results derived above are robust to this extension of the model. We also show that
incumbents will enter more new opportunities and that this will crowd out new
entrants. The increased levels of entry by incumbents raise costs of entry to new firms,
lowering their numbers.
2.5 Predictions of the Model
Our model predicts how the probability of entry into patenting depends on opportunity,
complexity, hold-up potential and incumbents’ experience. Here we summarize these
predictions, which are tested empirically below:10
Prediction 1: Greater technological opportunity increases the probability of entry.
Greater technological opportunity reduces competition for facets per opportunity,
which raises expected profits and thereby attracts entry.
Prediction 2: Greater complexity of a technology increases the probability of entry.
Greater complexity has countervailing effects: it reduces competition per facet as well
as duplicative R&D, attracting entry. It also increases the likelihood that some of a
technology remains unpatented, reducing its overall value and entry. Our model shows
that overall complexity increases entry.
Prediction 3: Greater potential for hold-up reduces the probability of entry.
Hold-up potential increases expected costs of entry, reducing it.
Prediction 4: More experienced incumbents are more likely to enter technological
opportunities new to them.
We show that incumbency advantage raises the number of opportunities that
incumbents enter. This implies that they also enter new opportunities, which they have
not previously been active in. This expansion of activity by incumbents crowds out
entry by new entrants.
3 Data and Empirical Model
This section of the paper describes the data we use in the empirical test of our
theoretical predictions. In particular, we discuss how we measure entry, how the set of
potential entrants is identified, and which measures and covariates are used.
Our empirical model is a hazard rate model of firm entry into patenting in a technology
10 Note that von Graevenitz et al. (2013) tested predictions from a substantially more restrictive version
of the model on the level of patent applications using data from the European Patent Office.
0
11
area as a function of technological opportunity, technological complexity, and hold-up
potential that characterize a technology area. Additional firm level covariates include
the age and size of firms. The models we estimate are stratified at the industry level.
That is, the unit of observation for each entry hazard is a firm-technology area, but the
hazard shapes and levels are allowed to vary by the industry that the firm is in. This
approach recognizes that patenting propensities vary across industries for reasons that
may not be technological (e.g., strategic reasons, or reasons arising from the historical
development of the sector).
We use a combination of firm level data for the entire population of UK firms registered
with Companies House and data on patenting at the European Patent Office and at the
Intellectual Property Office for the UK. The firm data come from the data held at
Companies House provided by Bureau van Dijk in their FAME database. European
patent registers do not include reference numbers from company registers, nor does
Bureau van Dijk provide the identification numbers used by patent offices in Europe.
Linking the data from patent registers to firm register data requires matching of
applicant names in patent documents and firm names in firm registers. In our work
both a firm’s current and previous name(s) were used for matching in order to account
for changes in firm names. For more details on the matching of firm- and patent-level
data see Appendix C.
Economic studies of entry are frequently hampered by the problem of identifying the
correct set of potential entrants (Bresnahan and Reiss, 1991; Berry, 1992). In our case
this problem is slightly mitigated by the fact that one set of potential entrants into
patenting in a specific technology area consists of all those firms that currently patent in
other technology areas. We complement this group of firms with a set of comparable
firms from the population of UK firms that have not patented previously.
To construct the sample we deleted all firms from the data for which we have no size
measure, because of missing data on assets. We select previously non-patenting firms
from the population of all UK firms in two steps: 1) we delete all firms in industrial
sectors with little patenting (amounting to less than 2 per cent of all patenting); and 2)
we choose a sample of non-patenting firms that matches our sample of patenting firms
by industry, size class, and age class. In principle, this approach will result in an
endogenous (choice-based) sample. However our focus is on industry and technology
area level effects rather than firm-level effects. Therefore we do not expect the sampling
approach we adopt to introduce systematic biases into the estimates we report. We
provide a number of robustness checks to ensure that our results are stable. These
reveal that sample composition does not affect the key results we present below. All
estimates are based on data weighted by the probability that a firm is in our sample.11
11 To check this, we estimated the model with and without weights based on our sampling methodology
and find little difference in the results.
12
The sample that results from our selection criteria is a set of firms with non-missing
assets in manufacturing, oil and gas extraction and quarrying, construction, utilities,
trade, and selected business services including financial services that includes all
(approximately 10,000) firms applying for a patent at the EPO or UKIPO during the
2001-2009 period and another 10,000 firms that did not apply for a patent.
The definition of technology areas that we use is based on the 2008 version of the ISI-
OST-INPI technology classification (denoted TF34 classes) (Schmoch, 2009). The list is
shown in Table 1, along with the number of EPO and UKIPO patents applied for by UK
firms with priority dates between 2002 and 2009. A comparison of the frequency
distribution of patenting across the technology areas from the two patent offices shows
that firms are more likely to apply for patents in Chemicals at the EPO, while Electrical
and Mechanical Engineering predominate in the national patent data (see the bottom
panel in Table 1).
We treat entry into each technology area as a separate decision made by firms. More
than half of firms we observe patent in more than one area and 10 per cent patent in
more than four. From the 20,000 firms observed, each of which can potentially enter
into each one of the 34 technology areas, we obtain about 700,000 observations at risk.
We cluster the standard errors by firm, so our models are effectively firm random
effects models for entry into 34 technology areas. Allowing firm choices to vary by
technology area is sensible under the assumption that firms’ patenting strategies are
contingent upon technology and industry level factors and are not homogeneous across
technology areas. We confirmed the validity of this assumption through interviews with
leading UK patent attorneys.
There are some technology-industry combinations that do not occur, e.g. audio-visual
technology and the paper industry, telecommunications technology and the
pharmaceutical industry. In order to reduce the size of the sample, we drop all
technology-industry combinations for which Lybbert and Zolas (2014) find no patenting
in their data and for which there was no patenting by any UK firm from the relevant
industry in the corresponding technology category. This removes about 30 per cent of
observations from the data. We provide a robustness check for this procedure in Table
E-2 in the Appendix.
[Table 1 here]
3.1 Variables
Dependent Variable - Entry
The dependent variable is a dichotomous variable taking the value one if a firm has
entered a technology area k at time t and otherwise the value zero. Entry into a
technology area is measured by the first time a firm applies for a patent that is classified
in that technology area, dated by the priority year of the patent.
13
Technological opportunity
Our first prediction from the theoretical model is that there will be more entry in
technology areas with greater technological opportunity. Additional reasons that a
sector may have more or less patenting include sector “size” or “breadth” and the
propensity of firms to patent in particular technologies for strategic reasons or because
of varying patent effectiveness in protecting inventions. To control for both
technological opportunity and these other factors, we include the logarithm of the
aggregate EPO patent applications in the technology sector during the year. To capture
opportunity more specifically we also include the past 5-year growth rate in the non-
patent (scientific publication) references cited in patents in that technology class at the
EPO.12 We have found that the growth rate in non-patent references is a better predictor
of entry than the level of non-patent references, which has been used previously.
Presumably the growth rate is a better indicator because it captures new or expanded
technological opportunity.
Technology complexity
The second prediction of the theoretical model is that technological complexity
increases entry, other things equal. Our interpretation of complexity is that it implies
many interconnections between inventions in a particular field, rather than a series of
fairly isolated inventions that do not connect to each other. To construct such a measure,
we use the concept of network density applied to citations among all the patents that
have issued in the particular technology area during the 10 years prior to the date of
potential entry. We use citations at the U.S. patent office, both because these are richer
(averaging 7 or so cites per patent during this period versus 3 for the EPO) and also to
minimize correlation with the thickets measure, which is based on EPO data.13
The network density measure is computed as follows: in any year t, there are Nkt patents
that have been applied for in technology area k between 1975 and year t. Each of these
patents can cite any of the patents that were applied for earlier, which implies that the
maximum number of citations within the technology area is given by Nkt(Nkt-1)/2. We
count the actual number of citations made and normalize them by this quantity, scaling
the measure by one million for visibility, given its small size.
Patent Thickets
The third prediction of our model is that greater potential for hold-up reduces entry. We
measure the potential for hold-up in patent thickets using the triples count proposed by
12 See von Graevenitz et al. (2013) for a more extensive discussion of this variable in the literature.
13 It is important to emphasize that although patent offices cooperate and share search reports citations
listed on U.S. patents are largely proposed by the applicant, whilst the citations listed on EPO and IPO
patents are inserted by the examiner. This explains why the two measures are not highly correlated.
14
von Graevenitz et al. (2011). This is a narrower interpretation of this measure than in
von Graevenitz et al. (2013), where it was used as a proxy for complexity and hold-up
potential together.14 In contrast, our model separates the effect of previously existing
patent thickets on entry from that of technological complexity. The hold-up potential of
thickets is captured by measuring how often patent applicants in a technology area
simultaneously face patents that overlap with and block their own applications from
firms that also block each-other’s patent applications. This is more likely to happen in
complex technologies, but does not necessarily arise as a result of complexity. Hence we
use separate measures of complexity and hold-up potential.
The triples measure is a count of the number of fully connected triads on the set of
firms’ critical patent citations. At time t each unidirectional link between two firms A
and B corresponds to one or more critical references to firm A’s patents in the set of
patents applied for by firm B in the years t, t-1 and t-2. These critical references are
obtained from examiner search reports issued by the EPO and represent prior art that
calls in question novelty and/or the inventive step of the patent application under
examination.15 Triples are then formed by groups of three firms where each firm has at
least one patent that is cited as critical prior art in the search report for at least one
patent held by each of the other two firms. That is, in a triple, each firm holds patents
that potentially block the other firms’ patents creating mutually blocking triads.
We use the same measure of triples as Harhoff et al. (2015), which contains all triples in
each technology area. The citation data used is extracted from PATSTAT (October 2011
edition).16 We normalize the count of triples by aggregate patenting in the same sector,
so that the triples variable represents the intensity with which firms potentially hold
blocking patents on each other relative to aggregate patenting activity in the technology.
The triples measure has been used in a number of papers since it was suggested by von
Graevenitz et al. (2011). They show that counts of triples by technical area are
significantly higher for technologies classified as complex than for areas classified as
14 In Appendix D, we show that this confounded the separate effects of complexity and hold-up. Including
the measures of complexity and hold-up potential proposed here in their empirical model, we find that
the effects on patenting incentives predicted by our theoretical model for complexity (positive) and hold-
up potential (negative) apply.
15 These are the so-called X- and Y-references in EPO search reports. According to 9.2.1 of the EPO’s
Guidelines for Examination: “Category "Y" is applicable where a document is such that a claimed
invention cannot be considered to involve an inventive step when the document is combined with one or
more other documents of the same category, such combination being obvious to a person skilled in the
art. If a document explicitly refers to another document as providing more detailed information on
certain features and the combination of these documents is considered particularly relevant, the primary
document should be indicated by the letter "X".”
16 Triples data was kindly provided by Harhoff et al. (2015).
15
discrete by Cohen et al. (2000). Fischer and Henkel (2012) find that the measure
predicts patent acquisitions by Non-Practicing Entities. Graevenitz et al. (2013) use the
measure to study patenting incentives in patent thickets and Harhoff et al. (2015) show
that opposition to patent applications falls in patent thickets, particularly for patents of
those firms that are caught up in thickets.
As a robustness check, we have also explored the use of duples, i.e. the count of mutual
blocking relationships, to measure hold-up potential. Combining both measures in one
regression leads to thorny problems of interpretation. Taken alone the measure has
similar effects as the triples measure in this context.
[Table 2 here]
Covariates
It is well known that firm size and industry are important predictors of whether a firm
patents at all (Bound et al. 1984 for U.S. data). Hall et al. (2013) show this for UK
patenting during the period studied here. Therefore, in all of our regressions we control
for firm size, industrial sector, and year of observation. We include the logarithm of the
firm’s reported assets and a set of year dummies in all the regressions.17 To control for
industrial sector, we stratify by industry, which effectively means that each industry has
its own hazard function, which is shifted up or down by the other regressors.
We also expect the likelihood that a firm will enter a particular technology area to
depend on its prior patenting experience overall, as well as its age. Long-established
firms are less likely to be exploring new technology areas in which to compete. Thus we
include the logarithm of firm age and the logarithm of the stock of prior patents applied
for in any technology by the firm, lagged one year to avoid any endogeneity concerns.
The variables on firm size and patent stock also allow us to test Prediction 4 about the
effect of incumbency advantage on entry.
3.2 Descriptive Statistics
Our estimation sample contains about 20,000 firms and 700,000 firm-TF34 sector
combinations. During the 2002-2009 period there are about 10,000 entries into
patenting for the first time in a technology area by these firms. Table C-1 in the appendix
shows the distribution of the number of entries per firm: 2,531 enter one class, and the
rest enter more than one. Table C-2 shows the population of UK firms obtained from
FAME in our industries, together with the shares in each industry that have applied for a
UK or European patent during the 2001-2009 period. These shares range from over 10
per cent in Pharmaceuticals and R&D Services to less than 0.1 per cent in Construction,
17 The choice of assets as a size measure reflects the fact that it is the only size variable available for the
majority of the firms in the FAME dataset.
16
Oil and Gas Services, Real Estate, Law, and Accounting.
3.3 Empirical Model
We use hazard models to estimate the probability of entry into a technology area. The
models express the probability that a firm enters into patenting in a certain area
conditional on not having entered yet as a function of the firm’s characteristics and the
time since the firm was “at risk,” which is the time since the founding of the firm. In
some cases, our data do not go back as far as the founding date of the firm, and in these
cases the data are “left-censored.” When we do not observe the entry of the firm into a
particular technology sector by the last year (2009), the data is referred to as “right-
censored.”
In Appendix E, we discuss the choice of the survival models that we use for analysis,
how to interpret the results, and present some robustness checks. We estimate two
classes of failure or survival models: 1) proportional hazard, where the hazard of failure
over time has the same shape for all firms, but the overall level is proportional to an
index that depends on firm characteristics; and 2) accelerated failure time, where the
survival rate is accelerated or decelerated by the characteristics of the firm. In the body
of the paper we present results using the well-known Cox proportional hazards model
stratified by industry. The effect of the stratification is that we allow firms in each of the industries to have a different distribution of the time until entry into patenting conditional on the regressors. That is, each industry has its own “failure” time distribution, where failure is defined as entry into patenting in a technology area, but the level of this distribution is also modified by the firm’s size, aggregate patenting in the technology, network density, and the triples density.
Appendix Table E-1 shows exploratory regressions made using various survival models. None of the choices made large differences to the coefficients of interest, for instance results
from the accelerated failure time models were similar to those of the Cox proportional
hazards model, but the estimated effects are somewhat larger (shown in Table E-1).
As indicated earlier, our data for estimation are for the 2002-2009 period, but many
firms have been at risk of patenting for many years prior to that. The oldest firm in our
dataset was founded in 1856 and the average founding year was 1992. Because the EPO
was only founded in 1978, we chose to use that year as the earliest date any of our firms
is at risk of entering into patenting. That is, we defined the initial year as the maximum
of the founding year and 1978. Table E-2 in the appendix presents estimates of our
model using 1900 instead of 1978 as the earliest at risk year and finds little difference
in the estimates.18 We conclude that the precise assumption of the initial period is
18 The main difference is in the firm age coefficient. Because the models are nonlinear, this coefficient is
identified even in the presence of year dummies and vintage/cohort (which is implied by the survival
17
innocuous. Our assumption amounts to assuming that the shape of the hazard for firms
founded between 1856 and 1978 but otherwise identical is the same during the 2002-
2009 period.
4 Results
Our estimates of the model for entry into patenting are shown in Table 3. All
regressions control for size, age, and industry. Both size and age are strongly positively
associated with entry into patenting in a new technological area. Our indicator of
technological opportunity and technology class size, the log of current patent
applications in the technology class, is also positively associated with entry into that
class, as predicted by our model.
Column 3 of Table 3 contains the basic result from our data and estimation, which is
fully consistent with the predictions of our theoretical model: greater complexity as
measured by citation network density increases the probability of entry into a
technology area (Prediction 2), as does technological opportunity (Prediction 1),
measured both as prior patenting in the class and as growth in the relevant science
literature. Controlling for both technological opportunity and complexity, firms are
discouraged from entry into areas with a greater density of triple relationships among
existing firms (Prediction 3). We interpret this latter result as an indicator of the
discouraging effect of hold-up possibilities or the legal costs associated with negotiation
of rights or defense in the case of litigation.
We were concerned that our network density (complexity) and triples density (hold-up
potential) measures might be too closely related to convey separate information, but we
found that the raw correlation between these two variables was -0.001. To check for the
impact of potential correlation conditional on year, industry, and the other variables, in
columns 1 and 2 of Table 3 we included these two measures of complexity/thickets
separately and found that although the coefficients were very slightly lower in absolute
value, the results still hold, although it is clear that the aggregate class size is correlated
negatively with the triples density via the denominator of the density (compare the
change in the log (patents in class) coefficient between columns 1 and 2).
As we show in Appendix E, the estimated coefficients in the table are estimates of the
elasticity of the yearly hazard rate with respect to the variable, and do not depend on
the industry specific proportional hazard. A one standard deviation increase in the log
of network density is associated with a 32 percent increase in the hazard of entry
(0.13*2.78), while a one standard deviation in the log of triples density is associated
model formulation). However it will be highly sensitive to the assumptions about vintage due to the age-
year-cohort identity.
18
with a 20 percent decrease in the hazard of entry (0.14*1.44). Thus the differences
across these technology areas in the willingness of firms to enter them is substantial,
bearing in mind that the average probability of entry is only about 1.5 per cent in this
sample.
[Table 3 here]
There are fixed costs to patenting, and a firm may be more likely to enter into patenting
in a new area if it already patents in another area. To test this idea, in the fourth column
of Table 3, we add the logarithm of past patenting by the firm. Inline with Prediction 4,
firms with a greater prior patenting history are indeed more likely to enter a new
technology area – doubling a firm’s past patents leads to an almost 100% higher hazard
of entry.
In the last column we interact the log of assets with the log of patents, the log of
network density, the growth of non-patent literature, and the log of triples density to
see whether these effects vary by firm size. The results show that the network density
and technological opportunity effects decline slightly with firm size. The triples density
effect does not show any size relationship, suggesting that hold-up concerns affect firms
of all sizes proportionately. We show this graphically in Figure 1, which overlays the
coefficients as a function of firm size on the actual size distribution of our firms. From
the graph one can see that the impact of aggregate patenting in a sector is higher and
more variable than the impact of the network density, and that both fall to zero for the
largest firms. Growth in non-patent literature is positively associated with technology
entry for small firms, but negatively for large firms, suggesting the role played by the
smaller firms in newer technologies based on science. Large firms seem not to be as
active in these areas. Controlling for all these features of a technology, the impact of
triples density is uniformly negative across firm size, which contradicts the view that
the potential for hold-up discourages entry by smaller firms more than by larger firms.
4.1 Robustness
One concern we may have with the relationship between entry and the triples variable
is simultaneity. That is, technology areas with lots of entry may also be prone to a
higher triples density, just because of the entries. To address this possibility, we use the
aggregate form of our entry regression. For each year we regress the log of the number
of first time entries in each technology-industry sector combination on the
characteristics of the technology class together with industry and year dummies. As
instruments for the triples density, we use the median examination lag in the
technology for patents applied for 5 and 6 years prior to the current year, which is long
enough so that most of them will have been granted, rejected, or withdrawn. The idea is
that classes with long examination lags may also be those where it is more difficult to
assess patentability, leading to the hold-up potential captured by the triples proxy
variable. We find that the instrumental variables regression easily passes the
19
specification tests for under-, weak and over-identification, justifying our choice of
instruments.
Table 4 shows the results, both ordinary least squares and instrumental variables.19
Note that we do not expect results to be identical when comparing the aggregate
regressions to individual firm-level hazard rate estimations, as the functional forms of
the models differ. However, the results are similar in sign to those in column 3 of Table
3, with the exception of the technological opportunity coefficient (the past five-year
growth in non-patent-literature references) which has the opposite sign. For our
purposes, interest centers on the coefficient of triples density. The least squares
estimate of the elasticity is negative and implies a 12 per cent reduction in entry per
year when the triples density increases by one standard deviation. Instrumenting this
variable triples its coefficient, which suggests that our hazard rate estimates may by an
underestimate of the true impact of potential hold-up on entry.
Table E-2 in the appendix explores some variations of the sample used for estimation in
Table 3. Column 1 of Table E-2 is the same as column 4 of Table 3 for comparison. The
first change (column 2) was to add back all the technology-industry combinations
where Lybbert and Zolas (2012) find no patenting in their data and where there was no
entry by any UK firm from the relevant industry into that technology category. These
observations are about 20 per cent of the sample. The impact of network density on
entry is weaker, but the impact of triples density and the technological opportunity
variables is considerably stronger. That is, technology area-industry combinations with
no patenting are also those where the technology area displays low technological
opportunity.
Next we removed all the firms with assets greater than 12.5 million pounds, to check
whether large firms were responsible for our findings.20 This removed about 2 per cent
of the 20,000 firms. Column 3 of Table E-2 shows that the results do not change a great
deal, although they are somewhat stronger. In column 4, we removed the
telecommunications technology sector from the estimation, because it is such a large
triples outlier. Once again, there was little change to the estimates. The last column of
Table E-2 shows the results of defining the minimum entry year as 1900. With the
exception of firm age, the coefficients are nearly identical to those in column 1 of the
table.
19 We also estimated this model by LIML and GMM, with almost no change in the resulting coefficients
(not shown).
20 12.5 million pounds is a cutoff based on the definition of Small and Medium-sized Enterprises (SMEs)
as firms with fewer than 250 employees. We do not have employment for all our firms, so we assume that
assets are approximately 50 thousand pounds per employee in order to compute this measure. For small
firms only, this yields an assets cutoff of 2.5 million pounds.
20
5 Conclusion
Patent thickets arise for a multitude of reasons; they are mainly driven by an increase in
the number of patent filings and concomitant reductions in patent quality (that is, the
extent to which the patent satisfies the requirements of patentability) as well as
increased technological complexity and interdependence of technological components.
The theoretical analysis of patent thickets (Shapiro, 2001) and the qualitative evidence
provided by the FTC in a number of reports (FTC, 2003; 2011) suggest that thickets can
impose significant costs on some firms. The subsequent literature has focused on the
measurement of thickets (e.g. Graevenitz et al. 2011; Ziedonis, 2004) and has linked
thickets to changes in firms’ intellectual property strategies in a number of dimensions.
There is still a lack of evidence on the effect of patent thickets as well as their welfare
implications at the aggregate level.
The empirical analysis of the effects of patent thickets must contend with two
challenges: first, patent thickets have to be measured and secondly, effects of thickets
must be separated from effects of other factors that are correlated with the growth of
thickets, in particular technological complexity.
This paper confronts both challenges. We show that our empirical measure for the
density of thickets captures effects of patent thickets predicted by theory. We separate
the impact of patent thickets on entry from effects of technological opportunity and
complexity and show that thickets reduce entry into patenting. Controlling for
technological opportunity and complexity is important because both are correlated with
entry into patenting and the presence of thickets. It is also worth emphasizing that our
measure of thickets is purged of effects that are driven by patenting trends in particular
technologies. That is, our results are not due to the level of invention and technological
progress within a technology field.
Our results demonstrate that patent thickets significantly reduce entry into those
technology areas in which growing complexity and growing opportunity increase the
underlying demand for patent protection. These are the technology areas, which are
associated most with productivity growth in the knowledge economy. However, the
welfare consequences of our finding are unclear. Reduced entry into new technology
areas could be welfare-enhancing: As is well known from the industrial organization
literature, entry into a market may be excessive if entry creates negative externalities
for active firms, for instance due to business stealing. This is likely to be true of
patenting too. Furthermore, Arora et al. (2008) show that the patent premium does not
cover the costs of patenting for the average patent (except for pharmaceuticals). These
and related facts might lead one to conclude that lower entry into patenting is likely to
increase welfare and that thickets raise welfare by reducing entry.
In contrast, reduced entry into patenting in new technology areas may also be welfare-
reducing, for at least two reasons. First, there is the obvious argument that the benefits
from more innovation may exceed any business stealing costs (as has been shown
21
empirically in the past by others, e.g., Bloom et al. 2013), so that some desirable
innovation may be deterred by high entry costs. Even if this were not true, there is no
reason to believe that firms that do not enter into patenting due to thickets are those we
wish to deter. Given the incumbency advantage, it is likely that the failure to enter into
patenting in these areas reflects less innovation by those who bring the most original
ideas, that is, by those who are inventing “outside the box.”
References
Allison, J.R., M.A. Lemley, and D. Schwartz, 2015. Our Divided Patent System. The
University of Chicago Law Review 82(3): 1073-1154.
Amir, R. and V. E. Lambson, 2000. On the effects of entry in Cournot markets. The Review
of Economic Studies 67(2): 235–254.
Arora, A., M. Ceccagnoli, and W.M. Cohen, 2008. R&D and the Patent Premium.
International Journal of Industrial Organization, 26(5): 1153-1179.
Berry, S., 1992. Estimation of a model of entry in the airline industry, Econometrica 60:
889-917.
Bessen, J., and E. Maskin, 2009. Sequential Innovation, Patents, and Imitation. RAND
Journal of Economics 40: 611-635.
Bessen, J. M. Meurer, and J. Ford, 2011. The Private and Social Costs of Patent Trolls.
Boston, MA: Boston University School of Law, Law and Economics Research Paper
No. 11-45.
Bessen, J. and M. Meurer, 2013. The Patent Litigation Explosion. Loyola University
Chicago Law Journal 45(2): 401-440.
Bessen, J. and M. Meurer, 2008. Patent failure: how judges, bureaucrats, and lawyers put
innovators at risk . Princeton, NJ: Princeton University Press.
Bloom, N., M. Schankerman, and J. van Reenen, 2013. Idenitfying technology spillovers
and product market rivalry, Econometrica 81 (4): 1347–1393.
Bound, J., C. Cummins, Z. Griliches, B.H. Hall, and A. Jaffe, 1984. Who does R&D and who
patents? R&D Patents and productivity, Z. Griliches (ed.). Chicago, University of
Chicago Press.
Bresnahan, T. and P. Reiss, 1991. Entry and competition in concentrated markets,
Journal of Political Economy 99: 977-1009.
Chien, C. and M. A. Lemley, 2012. Patent Holdup, the ITC, and the Public Interest. Cornell
Law Review 98(1).
Clark, D.J. and K. Konrad, 2008. Fragmented Property Rights and Incentives for R&D,
Management Science 54(5): 969-981.
22
Cockburn, I.M. and M.J. MacGarvie, 2011. Entry and Patenting in the Software Industry.
Management Science 57: 915–933.
Cohen, W.M., R.R. Nelson, and J.P. Walsh, 2000. Protecting their Intellectual Assets:
Appropriability Conditions and Why U.S. Manufacturing Firms Patent (or Not).
Cambridge, MA: NBER Working Paper No. 7552.
Farrell, J. and C. Shapiro, 2008. How strong are weak patents? American Economic
Review 98: 1347–1369.
Federal Trade Commission (FTC), 2003. To Promote Innovation - The Proper Balance of
Competition and Patent Law and Policy. Washington, DC: GPO. Available at
http://www.ftc.gov/reports/innovation/P040101PromotingInnovationandCompet
itionrpt0704.pdf
Federal Trade Commission (FTC), 2011. The Evolving IP Marketplace:. Aligning Patent
Notice and Remedies With Competition. Washington, DC. Available at
http://www.ftc.gov/os/2011/03/110307patentreport.pdf
Fershtman, C. and M. I. Kamien, 1992. Cross licensing of complementary technologies
International Journal of Industrial Organization 10(3): 329-348.
Fink C., M. Khan, and H. Zhou (2013). Exploring the worldwide patent surge. WIPO
Economic Research Working Paper No. 12.
Fischer, T. and J. Henkel, 2012. Patent trolls on markets for technology–An empirical
analysis of NPEs’ patent acquisitions. Research Policy, 41(9): 1519-1533.
Goolsbee, A. (1998). Does Government R&D Policy Mainly Benefit Scientists and
Engineers? The American Economic Review, 88(2): 298-302.
Graevenitz, von G., S. Wagner, and D. Harhoff, 2011. How to measure patent thickets—A
novel approach. Economics Letters 111: 1–4.
Graevenitz, von G., S. Wagner, and D. Harhoff 2013. Incidence and growth of patent
thickets: the impact of technological opportunities and complexity. Journal of
Industrial Economics, 61(3): 521-563.
Hall, B.H., C. Helmers, M. Rogers, and V. Sena, 2013. The importance (or not) of patents
to UK firms. Oxford Economic Papers 65 (3): 603-629.
Harhoff, D., G. von Graevenitz, and S. Wagner, 2015. Conflict Resolution, Public Goods
and Patent Thickets. Management Science, forthcoming.
Henderson, R., 1993. Underinvestment and Incompetence as Responses to Radical
Innovation: Evidence from the Photolithographic Alignment Equipment Industry,
The RAND Journal of Economics 24(2): 248-270.
Jaffe, A. and J. Lerner, 2004. Innovation and Its Discontents. Princeton University Press.
Joshi, A.M. and A. Nerkar, 2001. When do strategic alliances inhibit innovation by firms?
Evidence from patent pools in the global optical disc industry. Strategic
23
Management Journal 32(11): 1139-1160.
Lerner, J., 1995. Patenting in the shadow of competitors. Journal of Law and Economics
38: 463-495.
Lerner, J. and J. Tirole, 2004. Efficient Patent Pools. American Economic Review 94(3):
691-711.
Lerner, J., M. Strojwas, and J. Tirole. 2007. The design of patent pools: the determinants
of licensing rules, The RAND Journal of Economics 38(3): 610-625.
Lerner, J. and J. Tirole, 2013. Standard-Essential Patents, NBER Working Paper 19664.
Lewis, J.I.D. and R.M. Mott, 2013. The sky is not falling: Navigating the smartphone
patent thicket, WIPO Magazine, February 2013,
http://www.wipo.int/wipo_magazine/en/2013/01/article_0002.html
Lybbert, T. J. and N. J. Zolas, 2014. Getting Patents & Economic Data to Speak to Each
Other: An ‘Algorithmic Links with Probabilities’ Approach for Joint Analyses of
Patenting & Economic Activity. Research Policy 43(3): 530-542.
Milgrom P. and J. Roberts, 1994. Comparing equilibria. American Economic Review
84(3): 441–459.
Mulligan C. and T. B. Lee, 2012. Scaling the patent system. NYU Annual Survey of
American Law 68: 289-317.
Schmoch, U., 2009. Document IPC/CE/41/5, Annex, Conception of a Technology
Classification for Country Comparisons, 41th session of the IPC Committee of Experts.
1–15.
Shapiro, C., 2001. Navigating the patent thicket: Cross licenses, patent pools, and
standard setting. Innovation policy and the economy 1: 119–150.
Teece, D., E. Sherry, and P. Grindley, 2014. Patents and "Patent Wars" in Wireless
Communications: An Economic Assessment, Digiworld Economic Journal 95: 85-98.
U.S. Department of Justice (USDoJ), Federal Trade Commission (FTC), 2007. Antitrust
Enforcement and Intellectual Property Rights: Promoting Innovation and Competition,
1–217.
Vives, X., 2005. Complementarities and games: New developments. Journal of Economic
Literature XLIII(2): 437–479.
Ziedonis, R.H., 2004. Don't Fence Me In: Fragmented Markets for Technology and the
Patent Acquisition Strategies of Firms. Management Science 50: 804–820.
Technology categories GB pats EP pats Total GB pats EP pats
Elec machinery, energy 1,321 1,101 2,422 6.1% 4.4%
Audio‐visual tech 633 549 1,182 2.9% 2.2%
Telecommunications 1,181 1,206 2,386 5.5% 4.8%
Digital communication 590 732 1,323 2.7% 2.9%
Basic comm processes 302 146 447 1.4% 0.6%
Computer technology 1,481 1,302 2,783 6.8% 5.2%
IT methods for mgt 256 224 480 1.2% 0.9%
Semiconductors 269 248 518 1.2% 1.0%
Optics 392 481 873 1.8% 1.9%
Measurement 1,216 1,458 2,674 5.6% 5.8%
Analysis bio materials 132 426 557 0.6% 1.7%
Control 592 542 1,134 2.7% 2.2%
Medical technology 996 1,561 2,558 4.6% 6.3%
Organic fine chemistry 182 1,538 1,720 0.8% 6.2%
Biotechnology 193 950 1,143 0.9% 3.8%
Pharmaceuticals 277 1,876 2,153 1.3% 7.5%
Polymers 114 280 394 0.5% 1.1%
Food chemistry 88 458 547 0.4% 1.8%
Basic materials chemistry 314 1,050 1,363 1.5% 4.2%
Materials metallurgy 161 318 479 0.7% 1.3%
Surface tech coating 287 284 571 1.3% 1.1%
Chemical engineering 507 724 1,231 2.3% 2.9%
Environmental tech 296 344 640 1.4% 1.4%
Handling 996 813 1,809 4.6% 3.3%
Machine tools 428 356 784 2.0% 1.4%
Engines,pumps,turbine 887 942 1,829 4.1% 3.8%
Textile and paper mach 235 304 539 1.1% 1.2%
Other spec machines 742 623 1,365 3.4% 2.5%
Thermal process and app 410 261 671 1.9% 1.0%
Mechanical elements 1,149 854 2,002 5.3% 3.4%
Transport 1,063 930 1,993 4.9% 3.7%
Furniture, games 1,064 612 1,675 4.9% 2.5%
Other consumer goods 630 507 1,137 2.9% 2.0%
Civil engineering 2,237 960 3,196 10.3% 3.8%
Total 21,619 24,959 46,578
Electrical engineering 6,032 5,508 11,540 27.9% 22.1%
Instruments 3,328 4,468 7,796 15.4% 17.9%
Chemistry 2,418 7,822 10,240 11.2% 31.3%
Mechanical engineering 5,910 5,083 10,993 27.3% 20.4%
Other Fields 3,930 2,079 6,009 18.2% 8.3%
* Weighting by owners does not affect the numbers, since they all get added back into the same cell.
Weighting by classes means that a patent in multiple TF34 sectors is downweighted in each of the sectors.
Table 1
Patenting by Fame firms on Patstat (priority years 2002‐2009)
Weighted by #owners & #classes* Sector shares
Technology categories
Aggregate
EPO patents
Number of
EPO triples@
Triples per
1000 patents
US Citation
network
density#
Average non‐
patent
references
Elec machinery, energy 56,714 7751 136.7 39.4 0.420
Audio‐visual tech 34,131 13268 388.7 63.5 0.449
Telecommunications 62,288 27049 434.3 79.1 1.235
Digital communication 36,975 16529 447.0 178.5 1.397
Basic comm processes 10,035 2289 228.1 110.8 1.162
Computer technology 60,577 21956 362.4 54.3 1.529
IT methods for mgt 9,312 34 3.7 144.2 0.920
Semiconductors 24,544 9974 406.4 94.0 1.070
Optics 28,458 7767 272.9 58.0 0.806
Measurement 44,320 2503 56.5 45.6 1.000
Analysis bio materials 11,787 26 2.2 319.4 7.040
Control 17,612 308 17.5 112.1 0.445
Medical technology 66,062 4411 66.8 206.2 0.614
Organic fine chemistry 41,137 3993 97.1 32.9 5.253
Biotechnology 33,192 365 11.0 89.6 17.332
Pharmaceuticals 52,671 11222 213.1 76.6 6.391
Macromolecular chemistry 21,307 3722 174.7 92.7 1.236
Food chemistry 9,955 140 14.1 326.3 2.701
Basic materials chemistry 27,679 1929 69.7 84.8 1.498
Materials metallurgy 16,935 405 23.9 91.7 1.130
Surface tech coating 17,429 363 20.8 59.4 0.803
Chemical engineering 24,494 443 18.1 66.2 0.797
Environmental tech 12,708 858 67.5 206.9 0.487
Handling 30,343 252 8.3 66.9 0.137
Machine tools 24,040 508 21.1 64.0 0.191
Engines,pumps,turbine 32,602 6678 204.8 85.4 0.210
Textile and paper mach 23,145 2640 114.1 84.9 0.312
Other spec machines 29,826 319 10.7 65.7 0.422
Thermal process and app 15,290 335 21.9 146.3 0.189
Mechanical elements 32,716 1301 39.8 57.8 0.168
Transport 48,875 10929 223.6 68.3 0.203
Furniture, games 19,847 206 10.4 107.6 0.166
Other consumer goods 19,734 301 15.3 105.6 0.194
Civil engineering 28,817 171 5.9 117.1 0.150
Total 1,025,555 160,945 156.9 100.3 #REF!
Electrical engineering 294,575 98,850 335.6 60.4 1.042
Instruments 168,239 15,015 89.2 96.3 1.181
Chemistry 257,507 23,440 91.0 71.1 4.977
Mechanical engineering 236,836 22,962 97.0 70.1 0.227
Other Fields 68,398 678 9.9 110.8 0.167
@ Triples based on all EPO patenting, priority years 2002‐2009 (see text for definition and further explanation).
Table 2
UKIPO and EPO patents: numbers, triples and network density 2002‐2009
# Network density is 1,000,000 times the number of within technology citations between 1976 and the
current year divided by the potential number of such citations.
Variable
Log (network density) 0.115*** 0.127*** 0.107*** 0.184***
(0.024) (0.023) (0.021) (0.052)
Log (triples density ‐0.138*** ‐0.139*** ‐0.101*** ‐0.100***
in class) (0.011) (0.011) (0.010) (0.023)
Log (patents in class) 0.317*** 0.506*** 0.545*** 0.514*** 0.822***
(0.025) (0.031) (0.030) (0.027) (0.071)
5‐year growth of non‐ 0.060*** 0.084*** 0.072*** ‐0.009 0.103*
patent refs in class) (0.022) (0.022) (0.022) (0.021) (0.056)
Log assets 0.270*** 0.270*** 0.270*** 0.142*** 0.513***
(0.011) (0.011) (0.011) (0.013) (0.083)
Log firm age in years 1.135*** 1.135*** 1.136*** 0.773*** 0.767***
(0.104) (0.104) (0.104) (0.130) (0.131)
Log (pats applied for 0.836*** 0.836***
by firm previously) (0.021) (0.021)
Log (network density) ‐0.010*
* Log assets (0.006)
Log (triples density) ‐0.001
* Log assets (0.003)
Log (patents in class) ‐0.040***
* Log assets (0.008)
Log (average NPL refs) ‐0.015**
* Log assets (0.006)
Industry dummies stratified# stratified# stratified# stratified# stratified#
Year dummies yes yes yes yes yes
Log likelihood ‐65.96 ‐65.86 ‐65.84 ‐58.69 ‐58.67
Degrees of freedom 12 12 13 14 18
Chi‐squared 1270.6 1429.1 1517.2 3465.1 3458.6
The sample is matched on size class, sector, and age class. Estimates are weighted by sampling probability.
Coefficients for the hazard of entry into a patenting class are shown.
Standard errors are clustered on firm. *** (**) denote significance at the 1% (5%) level.
# Estimates are stratified by industry ‐ each industry has its own baseline hazard.
Table 3
Hazard of entry into patenting in a TF34 Class538,452 firm‐TF34 observations with 10,665 entries (20,384 firms)
Cox Proportional Hazard Model
Time period is 2002‐2009 and minimum entry year is 1978. Sample is UK firms with nonmissing assets, all patenting firms
and a matched sample of non‐patenting firms
Dependent variable
Coef. s.e.* Coef. s.e.*
Log (US network density) 0.018 0.029 0.055 0.032 *
Log (triples density) ‐0.079 0.010 *** ‐0.251 0.042 ***
Log (patent apps in class) 0.247 0.030 *** 0.469 0.059 ***
Past 5 year growth in NPL refs ‐0.099 0.015 *** ‐0.101 0.016 ***
R‐squared
Standard error
*Standard errors are clustered on class and industrial sector (which allows free correlation over time).
# Instruments are lag 5 and 6 median exam duration for patents in the class.
Log of triples density is treated as endogenous in the IV and LIML estimates
Table 4
0.564 0.608
Aggregate regressions for entry into patenting classes 2001‐20099 years*33 tech classes*26 sectors = 7722 observations
Log number of first time entries by a firm into class by sector
OLS IV#
0.478 0.389
0
2000
4000
6000
8000
10000
12000
14000
16000
18000
20000
‐0.4
‐0.2
0.0
0.2
0.4
0.6
0.8
1.0
Number o
f firms
Coefficient
Firm assets (millions of GBP)
Figure 1
Number of firms Aggregate patenting coefficient Network density coefficient
Triples density coefficient 5‐year growth in NPL refs
Appendix
A Previous Results
To help the reader this appendix summarizes a number of results derived by Graevenitz et al. (2013) aswell as some additional results that are useful.
A.1 The Probability of Patenting a Facet
The probability pk that a firm obtains a patent on a facet is:
pk(f 6k, F,NO(O,o 6k, N)) =
NO∑j=0
1
j + 1
(NO
j
)NO−j∏l=0
(1− fl
Fk
) NO∏m=NO−j
fmFk
. (A.1)
Then, the expected number of patents a firm owns when it applyies for fi facets is γk ≡ pkfi.For the comparative statics of entry stage it is useful to know that the elasticity of pk w.r.t. F is negative
if φk < 12:
∂pk∂Fk
=
NO∑i=0
1
i+ 1
(NO
i
)(1− φk)NO−iφik(−1)
(NO
Fk− NO − i
F − f
)(A.2)
Then the elasticity εpk,Fk is:
εpk,Fk =N2O
(φk − 1
2(1 + 1
NO))
1− φk(A.3)
A.2 The Expected Number of Rival Investors
The expected number of rival firms NO that undertake R&D on the same technology opportunity as firm i
can be expressed as a sum of products:
NO =N∑j=0
j
(N
j
)N−j∏l=0
(1− ωl)N∏
m=N−i
ωm. (A.4)
Graevenitz et al. (2013) show that NO is increasing in ωn, where n ∈ {l,m}.In the second stage equilibrium NO can be rewritten as:
NO =N∑j=0
j
(N
j
)(1− ωk)(N−j)ωjk. (A.5)
1
Incumbency Advantage
In the case in which there are incumbents and entrants the expected number of rival firms NO has torewritten slightly. To do this define:
ωIn ≡ oIi/O ωEn ≡ oEi
/O (A.6)
We assume that in a previous period Np firms entered and of these a fraction λ are still active. Thenthe expected number of rival firms NO that undertake R&D on the same technology opportunity as firm i
is:
NO =λNP∑j=0
j
(λNP
j
)(1− ωJk )(N−j)(ωJk )j +
N∑j=0
j
(N
j
)(1− ωEk )(N−j)(ωEk )j. (A.7)
A.3 The Expected Number of Facets Covered
In the second stage equilibrium the expected number of facets covered through the joint efforts of all firmsinvesting in a technological opportunity is:
Fk = F[1− (1− φk)(NO+1)
](A.8)
The derivative of this expression with respect to F is positive:
∂Fk∂Fk
= 1−(
1− φk)NO (
1 + φkNO
)≥ 0 . (A.9)
The elasticities of Fk with respect to fk and F are:
ηk =φk(1− φk)NO
1− (1− φk)(NO+1)(A.10)
εFk,Fk =1− (1− φk)(NO)(1 + φjNO)
1− (1− φk)(NO+1). (A.11)
which shows that 1 ≥ εFkFk ≥ 0 as the denominator in the fraction is always greater than the numerator. Itis useful to observe that the upper bound of the elasticity ηk is decreasing in NO. To see this note that theelasticity can be expressed as:
ηk =(1− φk)NO
(No + 1)(
1− φk No2!+ φ2
kNo(NO−1)
3!...) . (A.12)
This shows that the upper bound of the elasticity decreases in NO: limφk→0 ηk = 1/
(NO + 1) ≤ 1. Herewe use the binomial expansion of (1− φk)No+1. The expression also shows that the lower bound of ηk |φk=1
is zero.
2
B Results
This appendix contains derivations for the propositions set out in Section 2 of the paper.
B.1 Supermodularity of the Second Stage Game
This section sets out the main results needed to show that the second stage of game G∗ is supermodular.Consider the first order conditions that determine the equilibrium number of facets (f ) and technolog-
ical opportunities (o):
∂πik∂oi
= V∆(sik)− L(γik, sik)− Co(ΣNoj=1oj)− γikCa −
∂Cc∂oi
= 0 , (B.13)
∂πik∂fi
=oipk
Fk
([V µkηik
∆(sik)
sik− Fk
(∂L
∂γik+ Ca
)]+[Vd∆
dsik− ∂L
∂sik
](1− ηik)
)= 0 . (B.14)
Now, consider the cross-partial derivatives which must be positive, if the second stage game is supermod-ular. First, we derive the cross partial derivative with respect to firms’ own actions:
∂2πik∂oi∂fi
=pk
Fk
([V µkηik
∆(sik)
sik− Fk
(∂L
∂γik+ Ca
)]+[Vd∆
dsik− ∂L
∂sik
](1− ηik)
)= 0 . (B.15)
This expression corresponds to the first order condition (B.14) for the optimal number of facets.Now consider effects of rivals’ actions on firms’ own actions:
∂2πk∂oi∂om
=∂Fk∂om
sik
Fk
[V
∆
sik(µk − ξik) +
∂L
∂sik
]+∂pk∂om
fi
Fk
[(Vd∆
dsik− ∂L
∂sik
)− Fk
(∂L
∂γik+ Ca
)](B.16)
− ∂Co
∂ΣNoj=1oj
,
∂2πk∂oi∂fm
=∂Fk∂fm
sik
Fk
[V
∆
sik(µk − ξik) +
∂L
∂sik
]+∂pk∂fm
fi
Fk
[(Vd∆
dsik− ∂L
∂sik
)− Fk
(∂L
∂γik+ Ca
)],
(B.17)
∂2πk∂fi∂om
=∂Fk∂om
[∂V
∂Fk+∂2V
∂Fk2 Fkηik −
∂L
∂γik− Ca +
∂2L
∂sik2
sik
Fk(1− ηik)
]
+∂ηik∂om
(V
∆
sik(µk − ξik) +
∂L
∂sik
)− ∂pk∂om
[∂2L
∂γik2fi +
∂2L
∂sik2
fi
Fk(1− ηik)
], (B.18)
∂2πk∂fi∂fm
=∂Fk∂fm
[∂V
∂Fk
∆
sik
(ξik(1− εFk,f ) + εFk,f
)+∂2V
∂Fk2 Fkηik
∆
sik− ∂L
∂γik− Ca +
∂2L
∂sik2
sik
Fk(1− ηik)
]
+∂ηik∂fm
(V
∆
sik(µk − ξik) +
∂L
∂sik
)− ∂pk∂fm
[∂2L
∂γik2fi +
∂2L
∂sik2
fi
Fk(1− ηik)
]. (B.19)
The second stage game is supermodular, if the equations (B.16)-(B.19) are non-negative. The follow-
3
ing results show that the conditions noted in Section 2 above must hold simultaneously if the game issupermodular.
Using the first order condition (B.14), which will hold for any interior equilibrium, it can be shownthat: [(
Vd∆
dsik− ∂L
∂sik
)− Fk
(∂L
∂γik+ Ca
)]= −ηik
(V
∆
sik(µk − ξik) +
∂L
∂sik
). (B.20)
If(V ∆sik
(µk−ξik)+ ∂L∂sik
)> 0, then the second term in the cross-partial derivatives (B.16) and (B.17) is the
product of two negative expressions, and then equation (B.17) is positive. Equation (B.16) is also positivein a free entry equilibrium: the negative term at the end is less than the negative term in the derivative ofprofits w.r.t. Nk in Section 2, which is otherwise the same as equation (B.16): ∂Co
∂Nooo > ∂Co
∂Noo.
Turning to equations (B.18) and (B.19) we can show that:
∂ηik∂om
=∂2Fk∂fi∂om
fi
Fk− ∂Fk∂fi
∂Fk∂om
fi
Fk2 = −Fk
−1 ∂Fk∂om
(φk
1− φk+ ηik
)(B.21)
∂ηik∂fm
=∂2Fk∂fi∂fm
fi
Fk− ∂Fk∂fi
∂Fk∂fm
fi
Fk2 = −Fk
−1 ∂Fk∂fm
(φk
1− φk+ ηik
)(B.22)
This result allows us to rewrite equations (B.18) and (B.19) as follows:
∂2πk∂fi∂om
=1
Fk
∂Fk∂om
[(V
∆
sik(µk − ξik) +
∂L
∂sik
)(1− 2ηik −
φ
1− φ
)+∂2V
∂Fk2 Fkηik +
∂2L
∂sik2
sik
Fk(1− ηik)
]
− ∂pk∂om
[∂2L
∂γik2fi +
∂2L
∂sik2
fi
Fk(1− ηik)
], (B.23)
∂2πk∂fi∂fm
=1
Fk
∂Fk∂fm
[(V
∆
sik(µk − ξik) +
∂L
∂sik
)(1− 2ηik −
φ
1− φ
)+∂2V
∂Fk2 Fkηik +
∂2L
∂sik2
sik
Fk(1− ηik)
]
− ∂pk∂fm
[∂2L
∂γik2fi +
∂2L
∂sik2
fi
Fk(1− ηik)
]. (B.24)
Given assumptions (VF) and (LC) these two equations will be positive if(V ∆sik
(µk − ξik) + ∂L∂sik
)> 0 and(
1− 2ηik − φ1−φ
)> 0. We analyze each condition in more detail next:
1. Given our assumptions on the legal cost function (LC, eq. 4) the condition(V 4(sk)
sk(µk− ξk+ ∂L
∂sk
)>
0 ⇔ V 4(sk)sk
(µk − ξk > − ∂L∂sk
implies that µk > ξk. The elasticity of the value function w.r.t.additional covered patents must exceed the elasticity of the portfolio benefits function w.r.t. theshare of patents held by the firm. This condition is less restrictive than the assumption in Graevenitzet al. (2013) that µk > 1, since we are assuming that ξk < 1.
2. (1 − 2ηk) − φk1−φk
> 0 ⇔ (1 − 2φ) > (1 − φ)(NO+1). This holds for any φk < 12
and No suffi-ciently large. These restrictions imply a setting in which the ownership of patents belonging to eachopportunity is fragmented amongst many firms. It is more likely to arise if the technology is highlycomplex, otherwise the condition that φk < 1
2is less likely to hold.
4
In Appendix B.4 we derive the conditions under which the equilibrium of gameG∗ is unique. If there isa unique solution to the optimization problem of the firm at which profits are maximized, then this requiresthat ∂2πk/∂f
2 < 0. The restrictions that i) µk < 1 and ii) the share of overall profits which the firm obtainsis decreasing at the margin in the share of patents the firm holds (∂2∆/∂s2
k < 0) ensure that there is alwayssuch a unique interior solution.
In this game G∗ the comparative statics of patenting are the same as in the main model analyzed inGraevenitz et al. (2013). Specifically we can show that the following effects hold in this game:
∂2π
∂oi∂Fk> 0,
∂2π
∂fi∂Fk> 0,
∂2π
∂oi∂Ok
< 0,∂2π
∂fi∂Ok
< 0 (B.25)
This implies that complexity of the technology increases firms’ patent applications while increased tech-nological opportunity reduces firms’ patenting applications.
B.2 Effect of Hold-up on Patenting
Here we show that Proposition 2 holds. Consider the following cross-partial derivatives for the effects ofhigher legal costs L due to hold-up:
∂2πk∂o∂hk
= −∂L(γk, sk, hk)
∂hk< 0 , (B.26)
∂2πk
∂f∂hk= − opk
Fk
(Fk
(∂2L
∂γk∂hk
)+
∂2L
∂sk∂hk(1− ηk)
)< 0 . (B.27)
The first of these conditions shows that the expected legal costs of hold-up reduce the number of opportu-nities a firm invests in, in equilibrium. The second condition shows that firms with larger portfolios will bemore exposed to hold up and will benefit less from the share of patents they have patented per opportunity.Both of these effects reduce the number of facets each firm applies for.
B.3 Free Entry EquilibirumProposition 3There is a free entry equilibrium at which the marginal entrant can just break even, if R&D fixed costs peropportunity (Co) increase in the number of entrants.
In a free entry equilibrium it must be the case that the following conditions hold:
πk(ok, fk, Nk) > 0 ∧ πk(ok, fk, Nk + 1) < 0 . (B.28)
The effect of entry on profits at the first stage of game G∗ can be shown to be:
∂π(o, f)
∂Nk
= o∂NO
∂Nk
(sk
Fk
∂Fk∂NO
[V µ
∆(sk)
sk−(V (Fk)
∂∆
∂sk− ∂L
∂sk
)]
5
+∂pk∂NO
f
Fk
[(V (Fk)
∂∆
∂sk− ∂L
∂sk
)− Fk
(∂L
∂γk+ Ca
)]− ∂Co∂Noo
o
). (B.29)
This expression can be further simplified:
∂π(o, f)
∂Nk
= oεNO,NskN
((εFk,No − εpk,No ηk
) [V
∆(sk)
sk
(µk − ξk
)+∂L
∂sk
]− ∂Co
∂Noo
Noo
sk
). (B.30)
The first two terms in brackets in this derivative are positive and so is the third term. We can show thatthe limits of εFk,No and εFk,f in No are both zero. Therefore the above derivative is negative as long as theR&D fixed costs per opportunity are increasing in No. This is the condition set out in Proposition 3.
B.4 Uniqueness of the second stage equilibrium
We show that stage 2 of game G∗ is supermodular. This implies that there exists at least one equilibrium ofthe stage game. An alternative way of deriving existence of the second stage equilibrium for game G∗ is toanalyze the conditions under which the Hessian of second derivatives of the profit function (Hπ) is negativesemidefinite. This matrix consists of four derivatives of which only one leads to additional restrictions onthe model.
It is easy to see that ∂2π∂oi2
< 0 due to the coordination costs Cc(oi) and the restrictions we impose withassumption (FVC). The two cross-partial derivatives are both zero in equilibrium - refer to equation B.15.
Therefore, the only expression that remains to analyze is ∂2π∂fi
2 .
∂2π
∂fi2 =
oipk
Fk
[∂2V
∂Fk2
(∂Fk∂fi
)2∆Fkpk
+ 2∂V
∂Fk
∂Fk∂fi
d∆
dsik(1− ηik) + V
d2∆
s2ik
pk
Fk(1− ηik)2
− ∂2L
∂γik2p2k −
∂2L
∂sik2
(∂sik∂fi
)2
− 2
(Vd∆
dsik− ∂L
∂sik
)(1− ηik) ηik
fi
].
This can be further simplified:
∂2π
∂fi2 =
oipk
Fk
[∂2V
∂Fk2
(∂Fk∂fi
)2∆Fkpk
+ Vd2∆
s2ik
pk
Fk(1− ηik)2 − ∂2L
∂γik2p2k
− ∂2L
∂sik2
pk
Fk(1− ηik)2 − 2
(V
∆
sikξik(1− µk)−
∂L
∂sik
)(1− ηik) ηik
fi
]. (B.31)
If we impose the restriction that the second derivative of the value function is negative and that theelasticity of the value function, µk < 1, then the first and the last terms in the above expression arenegative. The sign of the second term in the expression depends on sign{ ∂2∆
∂sik2}, which we will assume is
negative. The third and fourth terms in the above expression are negative given the conditions imposed onthe legal cost function above.
6
B.5 Entry and Incumbency
In this section we analyze a game in which incumbents have lower costs of entry and demonstrate that ourmain predictions are robust.
We assume that a fraction λ (0 < λ < 1) of the previously active NP firms remain as incumbents. Thefirms enter until the marginal profit from entry is reduced to zero.
Objective Functions
First, consider the objective functions of incumbents and entrants and the patenting game they are involvedin. We analyze this game and show when it is supermodular.
Given symmetry of technological opportunities (Assumption S) the expected value of patenting forentrant and incumbent firm’s in a technology area k is:
πIik(oIi , f
Ii ) =oIi
V (Fk)∆(sIik)− L(γIik, sIik)−
Co(NPλ−1+NE∑j=1
oj)−Ψ
− f Ii pkCa− Cc(oIi ) .
(B.32)
πEik(oEi , f
Ei ) =oEi
V (Fk)∆(sEik)− L(γEik, sEik)− Co(
NPλ+NE−1∑j=1
oj)− fEi pkCa
− Cc(oEi ) . (B.33)
Define a game GE in which:
• There are NPλ incumbent firms and the number of entrants, NE , is determined by free entry.
• Entrants and incumbents simultaneously choose the number of technological opportunities oIi , oEi ∈
[0, On] and the number of facets applied for per opportunity f Ii , fEi ∈ [0, F n]. Firms’ strategy sets
Sn are elements of R4.
• Firms’ payoff functions πik, defined at (B.32,B.33), are twice continuously differentiable and dependonly on rivals’ aggregate strategies.
• Assumptions (VF, eqn. 1) and (LC, eqn. 4) describe how the expected value and the expected costof patenting depend on the number of facets owned per opportunity.
Firms’ payoffs depend on their rivals’ aggregate strategies because the probability of obtaining a patenton a given facet is a function of all rivals’ patent applications. Note that the game is symmetric withinthe two groups of firms as it is exchangeable in permutations of the players. This implies that symmetricequilibria exist, if the game can be shown to be supermodular (Vives, 2005).1
1Note also that only symmetric equilibria exist as the strategy spaces of players are completely ordered.
7
First order conditions for game GE:
∂πIik∂oIi
= V∆(sik)− L(γik, sik)−
Co(NPλ−1+NE∑j=1
oj)−Ψ
− γikCa − ∂Cc∂oIi
= 0 , (B.34)
∂πIik∂f Ii
=oIi pk
Fk
([V µεFkfi
4(sik)
sik− Fk
(∂L
∂γik+ Ca
)]+[V∂4(sik)
∂sik− ∂L
∂sik
](1− ηik)
)= 0 ,
(B.35)
∂πEik∂oEi
= V∆(sik)− L(γik, sik)− Co(NPλ+NE−1∑
j=1
oj)− γikCa −∂Cc∂oEi
= 0 , (B.36)
∂πEik∂fEi
=oEi pk
Fk
([V µεFkfi
4(sik)
sik− Fk
(∂L
∂γik+ Ca
)]+[V∂4(sik)
∂sik− ∂L
∂sik
](1− ηik)
)= 0 .
(B.37)
Proposition 7In game GE the equilibrium number of facets chosen by incumbents and entrants is the same: f I = fE .
We show in Appendix A.2 that in the game with incumbents the number of rivals per opportunity NO
becomes a function of both oI , oE . The first order conditions determining f I , fE both depend on the totalnumber of entrants per technological opportunity NO and so both on oI , oE . This is the only way in whichrivals’ choices of the number of opportunities to pursue enter these first order conditions2. Therefore thetwo conditions are identical and Proposition 7 holds.
Proposition 8The second stage of gameGE is smooth supermodular under the same conditions as gameG∗. Comparativestatics results for game G∗ also apply to game GE .
The first order conditions characterizing the game with incumbents and entrants are identical to thosefor the game without incumbents as long as Ψ = 0. As this variable is a constant it does not enterinto the second order conditions which we analyze to establish supermodularity and which underpin thecomparative statics predictions in Propositions 4-6.
Proposition 9In the second stage of game GE incumbents enter more technological opportunities, if they have a costadvantage in undertaking R&D (Ψ > 0).
The first order conditions determining the equilibrium number of opportunities chosen by incumbents andentrants are identical if firms R&D fixed costs per opportunity are the same (Ψ = 0). Therefore oI|Ψ=0 = oE .As the cost advantage of incumbents in undertaking R&D grows this increases the number of opportunitieschosen by incumbents:
∂2πIik∂oI∂Ψ
= 1 > 0 . (B.38)2Clearly the factors outside the brackets in equations (B.35), (B.37) also depend on these variables, but these do not affect
the equilibrium values of fEi , f I
j .
8
Proposition 10In the second stage of game GE the number of entrants decreases as the cost advantage of incumbentsincreases.
Due to the supermodularity of the second stage game, increases in incumbents’ choices of the numberof opportunities to invest in will raise the number of opportunities entrants invest in as well as the numbersof facets entrants and incumbents seek to patent in equilibrium. The increases in oI and oE will raise thefixed costs of entry into new opportunities, Co, which then reduces entry.
C Data
Our analysis relies on an updated version of the Oxford-Firm-Level-Database, which combines informationon patents (UK and EPO) with firm-level information obtained from Bureau van Dijks Financial AnalysisMade Easy (FAME) database (for more details see Helmers et al. (2011) from which the data descriptionin this section draws).
The integrated database consists of two components: a firm-level data set and IP data. The firm-leveldata is the FAME database that covers the entire population of registered UK firms. The original versionof the database, which formed the basis for the update carried out by the UKIPO, relied on two versions ofthe FAME database: FAME October 2005 and March 2009. The main motivation for using two differentversions of FAME is that FAME keeps details of inactive firms (see below) for a period of four years.If only the 2009 version of FAME were used, intellectual property could not be allocated to any firmthat has exited the market before 2005, which would bias the matching results. FAME is available since2000, which defines the earliest year for which the integrated data set can be constructed consistently.The update undertaken by the UKIPO used the April 2011 version of FAME. However, since there aresignificant reporting delays by companies, even using the FAME 2011 version means that the latest yearfor which firm-level data can be used reliably is 2009.
FAME contains basic information on all firms, such as name, registered address, firm type, industrycode, as well as entry and exit dates. Availability of financial information varies substantially across firms.In the UK, the smallest firms are legally required to report only very basic balance sheet information(shareholders’ funds and total assets). The largest firms provide a much broader range of profit and lossinformation, as well as detailed balance sheet data including overseas turnover. Lack of these kinds of datafor small and medium-sized firms means that our study focuses on total assets as a measure of firm sizeand growth.
The patent data come from the EPO Worldwide Patent Statistical Database (PATSTAT). Data on UKand EPO patent publications by British entities were downloaded from PATSTAT version April 2011.Due to the average 18 months delay between the filing and publication date of a patent, using the April2011 version means that the patent data are presumably only complete up to the third quarter in 2009.This effectively means that we can use the patent data only up to 2009 under the caveat that it might besomewhat incomplete for 2009. Patent data are allocated to firms by the year in which a firm applied forthe patent.
Since patent records do not include any kind of registered number of a company, it is not possible to
9
merge data sets using a unique firm identifier; instead, applicant names in the IP documents and firm namesin FAME have to be matched. Both a firm’s current and previous name(s) were used for matching in orderto account for changes in firm names. Matching on the basis of company names requires names in both datasets to be ‘standardized’ prior to the matching process in order to ensure that small (but often systematic)differences in the way names are recorded in the two data sets do not impede the correct matching. Formore details on the matching see Helmers et al. (2011).
Table C-1: Number of TF34 sectors entered between 2002 and 2009Number of sectors Number of firms Number of entries
1 2,531 2,531
2 1,347 2,694
3 647 1,941
4 271 1,084
5 155 775
6 71 426
7 45 315
8 29 232
9 20 180
10 14 140
11 4 44
12 2 24
13 3 39
14 0 0
15 or more 13 240
Total 5,152 10,665
10
Table C-2: Sample Population of Firms, by industry2-digit SIC3 Number of Number of Share patenting Number of
Industry firms patenters 2001-2009 patents
1 Basic metals 2,836 52 1.83% 231
2 Chemicals 3,834 246 6.42% 126
3 Electrical machinery 2,948 281 9.53% 727
4 Electronics & instruments 9,298 561 6.03% 444
5 Fabricated metals 24,681 606 2.46% 70
6 Food, beverage, & tobacco 8331 102 1.22% 29
7 Machinery 9,365 608 6.49% 313
8 Mining, oil&gas 83,491 15 0.02% 96
9 Motor vehicles 2,337 117 5.01% 22
10 Other manufacturing 94,952 1362 1.43% 150
11 Pharmaceuticals 1,008 105 10.42% 551
12 Rubber & plastics 6,094 398 6.53% 590
13 Construction 295596 372 0.13% 59
14 Other transport 3,292 89 2.70% 6274
15 Repairs & retail trade 128,266 251 0.20% 2324
16 Telecommunications 14,348 133 0.93% 2096
17 Transportation 60,837 75 0.12% 621
18 Utilities 12,880 75 0.58% 428
19 Wholesale trade 138,398 728 0.53% 3608
20 Business services 689,942 1639 0.24% 6757
21 Computer services 177,319 716 0.40% 1132
22 Financial services 183,042 219 0.12% 4930
23 Medicalservices 38,424 103 0.27% 1419
24 Personal services 94,791 196 0.21% 1194
25 R&D services 7,915 713 9.01% 168
26 inactive 37,525 271 0.72% 5673
Total 2,131,750 10,033 0.47% 40,032
11
Table C-3: Entry into techology area 2002-2009Numbers Shares
Total patenting in First time Pat’d previously Total First time Pat’d previously
Technology sector by GB firms patenter in another tech. entry patenter in another tech
Elec machinery, energy 1763.7 214 250 26.3% 12.1% 14.2%
Audio-visual tech 788.3 148 192 43.1% 18.8% 24.4%
Telecommunications 1874.4 146 179 17.3% 7.8% 9.5%
Digital communication 1054.0 92 141 22.1% 8.7% 13.4%
Basic comm processes 256.5 21 93 44.4% 8.2% 36.3%
Computer technology 2167.2 291 251 25.0% 13.4% 11.6%
IT methods for mgt 283.2 117 157 96.7% 41.3% 55.4%
Semiconductors 347.2 38 118 44.9% 10.9% 34.0%
Optics 584.7 55 130 31.6% 9.4% 22.2%
Measurement 1765.3 226 269 28.0% 12.8% 15.2%
Analysis bio materials 339.0 39 111 44.3% 11.5% 32.7%
Control 712.1 165 241 57.0% 23.2% 33.8%
Medical technology 1668.4 184 209 23.6% 11.0% 12.5%
Organic fine chemistry 1569.4 36 83 7.6% 2.3% 5.3%
Biotechnology 701.6 41 99 20.0% 5.8% 14.1%
Pharmaceuticals 1700.7 54 80 7.9% 3.2% 4.7%
Polymers 224.4 27 112 61.9% 12.0% 49.9%
Food chemistry 492.7 36 87 25.0% 7.3% 17.7%
Basic materials chemistry 1020.2 85 144 22.4% 8.3% 14.1%
Materials metallurgy 360.8 54 109 45.2% 15.0% 30.2%
Surface tech coating 400.7 77 195 67.9% 19.2% 48.7%
Chemical engineering 842.7 142 213 42.1% 16.9% 25.3%
Environmental tech 446.6 106 166 60.9% 23.7% 37.2%
Handling 1326.3 274 290 42.5% 20.7% 21.9%
Machine tools 577.8 106 180 49.5% 18.3% 31.2%
Engines,pumps,turbine 1443.4 82 160 16.8% 5.7% 11.1%
Textile and paper mach 442.0 77 137 48.4% 17.4% 31.0%
Other spec machines 847.4 180 234 48.9% 21.2% 27.6%
Thermal process and app 455.7 105 159 57.9% 23.0% 34.9%
Mechanical elements 1445.9 223 317 37.3% 15.4% 21.9%
Transport 1288.2 213 236 34.9% 16.5% 18.3%
Furniture, games 1239.2 288 223 41.2% 23.2% 18.0%
Other consumer goods 788.1 194 246 55.8% 24.6% 31.2%
Civil engineering 2045.8 463 255 35.1% 22.6% 12.5%
Total 33263.6 4599 6066 32.1% 13.8% 18.2%
12
D Robustness Tests on Network Density and Triples
This section examines how the addition of the network density measure to the analysis undertaken inGraevenitz et al. (2013) changes the coefficient and sign of the triples measure reported there. They exam-ine how complexity, technological opportunity and other variables affect the number of patents firms applyfor. Graevenitz et al. (2013) do not distinguish between hold-up and complexity in their model and use thetriples measure to capture complexity. The model presented in this paper separates the effects of hold-upand complexity and predicts that hold-up will reduce firms’ patenting incentives, while complexity raisesthese. In this paper, the network density measure is introduced as a measure of technological complexity,while we argue that the triples measure captures hold-up.
The exercise undertaken in this appendix is a validation of these two measures in light of the updatedmodel we present in this paper. The evidence provided is based on two data sets: first we report regressionresults obtained by adding the network density measure to the data used by (Graevenitz et al., 2013),second we report results obtained from a new dataset. This dataset covers the same period as that used by(Graevenitz et al., 2013), but it is based on the same more recent technology area classification as that usedin this paper. Furthermore, it is based on the same measure of triples as that used in this paper.
The models presented below are system GMM models which include a lagged dependent variable. Wedemonstrate that regardless of how the data are constructed the triples measure reduces patenting whilethe network density measure increases patenting efforts in our data. This supports our view that the triplesmeasure is a measure of hold-up.
In the results presented below we instrument potentially endogenous variables using lagged values. Ex-ogeneity of the instruments is tested using difference in Hansen tests. We instrument the lagged dependentvariable and its interaction with fourth order lags. All other variables are instrumented with third orderlags or higher. We include only year and area dummies in the levels equations as it is likely that the fixedeffects are correlated with differences in the remaining explanatory variables.
Instrument sets are collapsed in order to reduce the number of instruments used. Throughout we relyon the Hansen test to determine whether instruments are exogenous. Where the statistic indicated that thiswas not the case we rejected the models. We report only those models that were not rejected by the testfor which the lagged dependent variable was within the range one would expect from estimation of OLSmodels with the same specification.
D.1 Sample and Definition of Variables
The sample used for both tables below consists of all firms that have at least one hundred patent applicationsat EPO across all teachnology areas between 1987 and 2002 and who have applied for patents in at leastthree years in the sample period in a technology area.
The two tables below include a number of variables that we do not use in this paper other than here.We briefly discuss these variables next:
Dependent variable In both tables below the dependent variable is the logarithm of the number ofpatents each firm has applied for in a technology area and year. To deal with missing values arising
13
from firms not having patent applications in some years we add one to all patent counts before taking thelogarithm.
Triples count In Table D-1 below we use the triples count employed by Graevenitz et al. (2013). Theycount how often firm triples arise, such that each firm in a triple holds patents that are cited as limitingone or more patent applications submitted by each of the other two firms. Their measure of triples isconstructed using only the ten most frequently cited firms in each applicant’s patent portfolio in any areaand year. In Table D-2 we use the same triples count as in this paper, i.e. the restriction to the mostfrequently cited firms is removed.
Fragmentation The fragmentation measure used here is based on Ziedonis (2004). The measure is basedonly on critical references and captures the concentration of prior art cited in the patent portfolio of a firmin a year and area.
Large / Relative Size In Table D-1 below we use a dummy variable that is one for all firms above themedian firm by size of patent portfolio in each area and year. In Table D-2 we capture relative size bymeasuring the size of each firm’s patent portfolio by area up to a given year relative to the total number ofall firms’ patent applications in that area and year.
D.2 Data used by Graevenitz et al. (2013)
The results presented here are based on adding the network density measure of complexity constructed forthis paper from US patent data to the data used by Graevenitz et al. (2013). They study the determinantsof the level of patent applications at EPO.
Column E in Table D-1 below is reported by Graevenitz et al. (2013) and is presented here as a referencepoint. Models K,L,M contain new results. Model K replicates model E closely, differences are likely dueto updates to the code used to estimate these models. Model L is like model K with the network densitymeasure added. In model M we adjust the set of instruments to obtain a model with a lagged dependentvariable that is significantly below 1 at the 5% level at the mean of the triples variable.
These results show that adding the network density measure to the data does not change the signor significance of the other variables reported in Table D-1. Network density itself has a positive andsignificant effect on patenting in models L and M. We would expect to see this, if this measure capturescomplexity.
Adding network density does have an important and not immediately obvious effect. The range ofvalues of non-patent references for which an increase in the hold-up measure (triples) reduces patentingincentives is larger in model M than model E at the mean of the patent count: in model E non-patentreferences must lie beyond 1.22 for an increase in triples to have a negative effect on patenting incentives,while in model M non-patent references beyond 1.1 have the same effect. Similarly the range of valuesof the patent count for which an increase in the hold-up measure (triples) reduces patenting incentives islarger in model M than model E at the mean of non patent references. This shows that the triples measure
14
contained in this data has a negative effect on patenting incentives in sufficiently complex technologies.We would expect to see this, if triples is a measure of hold-up.
Table D-1: GMM Models for Patent Applications - Old DataVariable SGMM E SGMM K SGMM L SGMM Mlog Patentcountt−1 0.749∗∗∗ 0.879∗∗∗ 1.020∗∗∗ 0.976∗∗∗
(0.093) (0.099) (0.113) (0.106)
log Patentcountt−1× Triples −0.017∗∗∗ −0.017∗∗∗ −0.013∗∗∗ −0.012∗∗∗(0.003) (0.002) (0.002) (0.002)
Non Patent References (NPR) 1.553∗∗∗ 1.863∗∗∗ 1.648∗∗∗ 1.389∗∗∗(0.254) (0.252) (0.240) (0.183)
NPR × Triples −0.036∗∗∗ −0.038∗∗∗ −0.033∗∗∗ −0.028∗∗∗(0.006) (0.005) (0.005) (0.004)
NPR × Triples × Large 0.007∗∗∗ 0.006∗∗∗ 0.005∗∗∗ 0.005∗∗∗(0.002) (0.001) (0.001) (0.001)
NPR × Large −0.366∗∗∗ −0.386∗∗∗ −0.339∗∗∗ −0.340∗∗∗(0.081) (0.062) (0.051) (0.043)
Fragmentation −0.474∗∗ −0.521∗∗ −0.543∗∗ −0.490∗∗∗(0.170) (0.182) (0.174) (0.129)
Fragmentation × Triples 0.006 0.009∗ 0.005 0.004
(0.006) (0.004) (0.004) (0.004)
Triples 0.055∗∗∗ 0.052∗∗∗ 0.046∗∗∗ 0.039∗∗∗(0.010) (0.007) (0.007) (0.005)
Areas 0.096∗∗∗ 0.084∗∗∗ 0.049∗∗ 0.050∗∗(0.012) (0.010) (0.017) (0.016)
Large 0.342∗∗ 0.476∗∗∗ 0.427∗∗∗ 0.424∗∗∗(0.117) (0.105) (0.091) (0.072)
Network Density 0.002∗∗ 0.002∗∗(0.001) (0.001)
Year dummies YES YES YES YESPrimary area dummies YES YES YES YESConstant −1.443∗∗∗ −1.846∗∗∗ −2.017∗∗∗ −1.772∗∗∗
(0.319) (0.267) (0.270) (0.237)
N 173448 173448 173448 173448
m1 −10.860 −10.470 −9.318 −9.634
m2 4.739 6.307 6.2 6.305
m3 .896 .509 −.071 −.302
Hansen 10.988 2.017 4.454 12.506
p-value .052 .569 .216 .052
Degrees of freedom 5 3 3 6
* p<0.05, ** p<0.01, *** p<0.001
15
1. Asymptotic standard errors, asymptotically robust to heteroskedasticity are reported in parentheses2. m1-m3 are tests for first- to third-order serial correlation in the first differenced residuals.3. Hansen is a test of overidentifying restrictions. It is distributed as χ2 under the null of instrument
validity, with degrees of freedom reported below.4. In all cases GMM instrument sets were collapsed and lags were limited.
D.3 New Data at 34 Area Level
Here we present results based on an updated dataset of patenting in Europe that is based on PATSTAT,October 2014, but covers the same range of years (1987-2002), for better comparability with the datapresented in the previous section.
Table D-2: GMM Models for Patent Applications - New DataVariable SGMM E SGMM K SGMM L
log Patentcountt−1 0.454 ∗ ∗∗ 0.799 ∗ ∗∗ 0.808 ∗ ∗∗(0.111) (0.080) (0.082)
Non Patent References (NPR) −0.014 ∗ ∗∗ −0.014 ∗ ∗∗ −0.014 ∗ ∗∗(0.001) (0.001) (0.001)
Triples −0.091 ∗ ∗∗ −0.100 ∗ ∗∗ −0.117 ∗ ∗∗(0.016) (0.011) (0.012)
Network Density −0.216 0.195 0.217∗(0.129) (0.101) (0.107)
Fragmentation 0.899 ∗ ∗∗ 0.658 ∗ ∗∗ 0.804 ∗ ∗∗(0.155) (0.093) (0.085)
Areas 0.020∗ 0.016 ∗ ∗ −0.001
(0.009) (0.006) (0.006)
Relative Size −0.010 ∗ ∗∗ 0.000 −0.000
(0.003) (0.003) (0.002)
Constant −0.033 −0.208 ∗ ∗∗ −0.149 ∗ ∗(0.045) (0.052) (0.046)
N 168066 168066 168066
m1 −7.598 −12.948 −13.08
m2 3.871 9.964 9.896
m3 −1.450 .538 .055
Hansen 3.49 3.79 2.1
p-value .48 .29 .55
Degrees of freedom 4 3 3
* p<0.05, ** p<0.01, *** p<0.0011. Asymptotic standard errors, asymptotically robust to heteroskedasticity are reported in parentheses2. m1-m3 are tests for first- to third-order serial correlation in the first differenced residuals.3. Hansen is a test of overidentifying restrictions. It is distributed as χ2 under the null of instrument
16
validity, with degrees of freedom reported below.4. In all cases GMM instrument sets were collapsed and lags were limited.
The main difference between the data used here and the older data used by Graevenitz et al. (2013)is that we now rely on a more recent, slightly finer specification of the number of technology areas: thecurrent classification contains 34 rather than 30 areas. In addition, the triples measure we use now capturesall triples and not just those affecting each firm and its ten closest technology rivals as in the earlier data.Due to the larger number of areas we now exclude slightly more patentees when applying the criterion thata firm must have at least one hundred patents in a technology area and must have at least three years ofpatent activity in an area to be included in the analysis.
Table D-2 demonstrates that the predicted negative effects of triples and non-patent references arepresent in all specifications we report. We also show that network density is either not significant orpositive and significant. The model in which the measure is positive and significant is our prefered model,due to the low instrument count and the better test of overidentifying restrictions.
Overall, both sets of models demonstrate that an interpretation of the triples measure as a measure ofhold-up only, and of the network density as a measure of complexity, is consistent with the effects weobserve in the data on patent applications presented here.
E Estimating Survival Models
This appendix gives some further information about the various survival models we estimated and therobustness checks that were performed. We estimated two general classes of failure or survival models: 1)proportional hazard, where the hazard of failure over time has the same shape for all firms, but the overalllevel is proportional to an index that depends on firm characteristics; and 2) accelerated failure time, wherethe survival rate is accelerated or decelerated by the characteristics of the firm. We transform (2) to ahazard rate model for comparison with (1), using the usual identity between the probability of survival totime t and the probability of failure at t given survival to t− 1. The first model has the following form:
Pr (i first patents in j at t | i has no patents in j ∀s < t,Xi)
h(Xi, t) = h(t) exp (Xi, β)
where i denotes a firm, j denotes a technology sector, and t denotes the time since entry into the sample.h(t) is the baseline hazard, which is either a non-parametric or a parametric function of time since entryinto the sample. The impact of any characteristic x on the hazard can be computed as follows:
∂h(Xi, t)
∂xi= h(t) exp (Xi, β)β or
∂h(Xi, t)
∂xi
1
Xi, t= β (E.39)
Thus if x is measured in logs, β measures the elasticity of the hazard rate with respect to x. Note thatthis quantity does not depend on the baseline hazard h(t), but is the same for any t. We use two choices for
17
h(t): the semi-parametric Cox estimate and the Weibull distribution ptp−1. By allowing the Cox h(t) orp to vary freely across the industrial sectors, we can allow the shape of the hazard function to be differentfor different industries while retaining the proportionality assumption.
In order to allow even more flexibility across the different industrial sectors, we also use two acceleratedfailure time models, the log-normal model and the log-logistic model. These have the following basic form:
Log-normal :S(t) = 1− Φ
{ln (t)− µ
σ
}(E.40)
Log-logistic :S(t) =1
1 + (λt)1γ )
(E.41)
where S(t) is the survival function and λi = exp (Xiβ). We allow the parameters σ (log-normal) or γ(log-logistic) to vary freely across industries (j). That is, for these models, both the mean and the varianceof the survival distribution are specific to the 2-digit industry. In the case of these two models, the elasticityof the hazard with respect to a characteristic x depends on time and on the industry-specific parameter (σor γ), yielding a more flexible model. For example, the hazard rate for the log-logistic model is given bythe following expression:
h(t) =−d logS(t)
dt=
λ1γ t
1γ−1
γ(
1 + (λt)1γ )) (E.42)
From this we can derive the elasticity of the hazard rate with respect to a regressor x:4
∂ log hij(t)
∂xi=
−β(1 + (λt)
1γ
) (E.43)
One implication of this model is therefore that both the hazard and the elasticity of the hazard withrespect to the regressors depend on t, the time since the firm was at risk of patenting. We sample the firmsduring a single decade, the 2000s, but some of the firms have been in existence since the 19th century.This fact creates a bit of a problem for estimation, because there is no reason to think that the patentingenvironment has remained stable during that period. We explored variations in the assumed first date atrisk in Tables E-1(1978) and E-2 (1900), finding that the choice made little difference. Accordingly, wehave used a minimum at risk year of 1978 for estimation in the main table in the text.
4We assume that x is in logarithms, as is true for our key variables, so this can be interepreted as an elasticity.
18
Table E-1Hazard of entry into patenting in a TF34 Class - Comparing models
538,452 firm-TF34 observations with 10,665 entries (20,384 firms)
Proportional hazard AFTVariable Cox PH Weibull Log logistic Log normal
Log (network density) 0.108 ∗ ∗∗ 0.111 ∗ ∗∗ 0.308 ∗ ∗∗ 0.247 ∗ ∗∗(0.021) (0.021) (0.096) (0.040)
Log (triples density −0.100 ∗ ∗∗ −0.098 ∗ ∗∗ −0.511 ∗ ∗∗ −0.258 ∗ ∗∗in class) (0.010) (0.010) (0.071) (0.024)
Log (patents in class) 0.513 ∗ ∗∗ 0.513 ∗ ∗∗ 2.177 ∗ ∗∗ 1.095 ∗ ∗∗(0.027) (0.027) (0.304) (0.089)
5-year growth of non-patent (0.013) −0.001 −0.077 −0.039
refs in class (0.022) (0.021) (0.084) (0.038)
Log assets 0.149 ∗ ∗∗ 0.130 ∗ ∗∗ 0.529 ∗ ∗∗ 0.198 ∗ ∗∗(0.013) (0.013) (0.082) (0.024)
Log (pats applied for 0.848 ∗ ∗∗ 0.860 ∗ ∗∗ 5.685 ∗ ∗∗ 3.973 ∗ ∗∗by firm previously) (0.021) (0.019) (0.917) (0.368)
Industry dummies stratified stratified stratified stratifiedYear dummies yes yes yes yes
Log likelihood −58.8 −96, 051.0 −114, 127.0 −111, 662.1
Degrees of freedom 13 38 38 38
Chi-squared 3183.2 4560.1 168.4 300.1
All estimates are weighted estimates, weighted by sampling probability. For the Cox and Weibullmodels, coefficients shown are elasticities of the hazard w.r.t. the variable. For the log-logistic, -beta is shown.*** (**) denote significance at the 1% (5%) level.Time period is 2002-2009 and minimum entry year is 1978. Sample is all UK firms with nonmissing assets.AFT - Accelerated Failure Time models# Estimates are stratified by industry - each industry has its own baseline hazard.
19
Table E-2Hazard of entry into patenting in a TF34 Class - Robustness
Variable (1) (2) (3) (4) (5) Big firms
Log (network density) 0.107∗∗∗ −0.008 0.114∗∗∗ 0.107∗∗∗ 0.108∗∗∗ 0.390∗∗∗(0.021) (0.019) (0.024) (0.021) (0.021) (0.145)
Log (triples density −0.101∗∗∗ −0.183∗∗∗ −0.114∗∗∗ −0.103∗∗∗ −0.103∗∗∗ −0.099∗in class) (0.010) (0.008) (0.011) (0.009) (0.009) (0.055)
Log (patents in class) 0.514∗∗∗ 0.623∗∗∗ 0.605∗∗∗ 0.520∗∗∗ 0.518∗∗∗ 0.390∗∗(0.027) (0.024) (0.031) (0.027) (0.027) (0.157)
5-year growth of non- −0.009 −0.196∗∗∗ −0.002 −0.012 −0.012 −0.125
patent refs in class (0.021) (0.020) (0.025) (0.021) (0.021) (0.135)
Log assets 0.142∗∗∗ 0.138∗∗∗ 0.186∗∗∗ 0.139∗∗∗ 0.146∗∗∗ 0.303
(0.013) (0.013) (0.018) (0.013) (0.013) (0.233)
Log firm age in years 0.773∗∗∗ 0.588∗∗∗ 0.739∗∗∗ 0.778∗∗∗ 0.021 0.710
(0.130) (0.145) (0.155) (0.131) (0.246) (0.672)
Log (lagged firm-level) 0.836∗∗∗ 0.947∗∗∗ 0.954∗∗∗ 0.840∗∗∗ 0.878∗∗∗ 0.512∗∗∗patent stock (0.021) (0.019) (0.033) (0.021) (0.021) (0.132)
Industry dummies yes yes yes yes yes yes
Observations 538,452 692,038 452,313 523,547 538,452 5,655Firms 20,384 20,384 17,993 20,384 20,384 255Entries 10,665 10,665 8,149 10,340 10,665 219Entry rate 1.98% 1.54% 1.96% 1.97% 1.98% 3.87%
Log likelihood −58.69 −54.60 −40.59 −56.94 −53.81 −0.21
Degrees of freedom 14 14 14 14 14 14
Chi-squared 3465.1 5065.8 1688.8 3459.1 3137.4 51.3
All estimates are weighted estimates, weighted by sampling probability. Coefficients shown are negativeof the estimates (larger coefficient increases entry probability).*** (**) denote significance at the 1% (5%) level.Time period is 2002-2009. Sample in (1) is all UK firms with nonmissing assets.Log-logistic model stratified by industry.(1) Estimates from Table 3, for comparison.(2) Observations for tech sectors of firms whose industry has no such patenting (Lybbert and Zolas, 2014) andwhere there is no entry by any UK firm in that industry are dropped.(3) SMEs: firms with assets > 12.5 million GBP removed.(4) The Telecom tech sector is removed.(5) The minimum founding year is 1900 instead of 1978.
20
References
GRAEVENITZ, G., S. WAGNER, AND D. HARHOFF (2013): “Incidence and Growth of Patent Thickets: The Impact
of Technological Opportunities and Complexity,” The Journal of Industrial Economics, 61, 521–563.
HELMERS, C., M. ROGERS, AND P. SCHAUTSCHICK (2011): “Intellectual Property at the Firm-level in the UK:
The Oxford Firm-level Intellectual Property Database,” Tech. Rep. DP 546.
LYBBERT, T. J. AND N. J. ZOLAS (2014): “Getting Patents and Economic Data to Speak to each other: An Al-
gorithmic Links with Probabilities Approach for Joint Analyses of Patenting and Economic Activity,” Research
Policy, 43, 530–542.
VIVES, X. (2005): “Complementarities and Games: New Developments,” Journal of Economic Literature, XLIII,
437–479.
ZIEDONIS, R. H. (2004): “Don’t Fence Me In: Fragmented Markets for Technology and the Patent Acquisition
Strategies of Firms,” Management Science, 50, 804–820.
21
Top Related